Active damper impedance self-adaption method based on SOM neural network
By introducing the active damper impedance adaptation method of SOM neural network in the power system, the resonance problem caused by the change of the grid impedance is solved, the pure resistance characteristics and efficient resonance suppression at low switching frequency are realized, and the system stability is improved and power efficiency.
Patent Information
- Application Number
- CN202411073403.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-06
- Publication Date
- 2024-11-22
AI Technical Summary
In modern power systems, changes in power grid impedance lead to resonance problems in inverter systems, and the prior art is difficult to effectively suppress resonance, affecting system stability and performance.
Adoption method of active damper impedance based on SOM neural network is adopted, and the SOM neural network is introduced through the virtual resistor link, screening and clustering common point voltage resonance components, and the weight is updated in real time, so that the active damper presents in the target resonant frequency band. Pure resistance characteristics improve resonance suppression capability and reduce power loss.
It realizes that the output impedance of the active damper exhibits pure resistance characteristics at lower switching frequency, significantly improves the resonance suppression effect, reduces switching losses, and improves the equipment power, meeting the stability requirements within the wide band.
Smart Images

Figure CN119005076A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of power electronics, and in particular to an active damper impedance self-adaptation method based on a SOM neural network. Background Art
[0002] The operating characteristics of traditional power systems are determined by synchronous generators, and the output voltage of the generator can be considered as a sinusoidal voltage with extremely low harmonic components. When the penetration rate of power electronic converters is low, the interactive coupling between the power electronic converters and the power grid is relatively limited, and the power system can maintain characteristics close to that of an ideal voltage source. However, the large-scale use of power electronic converters with fast response speed and flexible control in modern power systems has made the "machine-grid" interaction increasingly frequent and the interaction mechanism increasingly complex. In the construction of new power systems, the operating state of power electronic converters has an increasingly significant impact on the stability and efficiency of the power grid. Due to the access of a large number of power electronic equipment and the existence of line impedance, the impact of grid impedance cannot be ignored. The grid-connected inverter is connected to the power system through a common point. As the grid impedance increases, the load current will introduce a variety of background harmonics into the grid voltage, breaking the preset ideal sinusoidal voltage source. On the other hand, when the grid impedance fluctuates over a wide range, it may also cause the inverter group to resonate over a wider frequency band. In modern power systems, it is particularly important to design an effective resonance suppression strategy to improve the stability of the inverter system under the condition of grid impedance changes. Summary of the invention
[0003] The purpose of the present invention is to provide an active damper impedance adaptive method based on a SOM neural network based on the characteristics of the prior art, introduce the SOM neural network in the virtual resistance link, take the common point voltage resonance component as input, judge the winning neuron through the winning neuron selection equation, the power of the active damper determines the screening limit, cluster the common point voltage resonance component according to the limit, add weights to the resonance component, and the weights change with the resonance size, so that the neighborhood neurons and the winning neurons update the weights in real time towards the optimal goal, so that the weight of the largest resonance component at each frequency is the largest, R v The value is determined according to the largest resonant component, that is, the active damper only exhibits resistance characteristics within the target resonant frequency band, thereby improving the resonance suppression capability and reducing power loss.
[0004] In order to achieve the above object, the present invention adopts the following technical solutions:
[0005] An active damper impedance adaptive method based on SOM neural network, comprising:
[0006] Step S1: the active damper damps the voltage resonance of the common point by constructing a virtual impedance, and the virtual impedance presents a resistance characteristic to the outside;
[0007] Step S2: Establish a SOM neural network consisting of an input layer and a competition layer, wherein the input layer receives the resonant component from the common point; the competition layer selects the resonant component greater than the screening limit according to the selection equation by means of analysis, comparison and discovery, and obtains a resonant frequency distribution with a high content;
[0008] Step S3: the active damper adds a filter in the virtual resistance loop to eliminate the interference of low-frequency harmonics on the function of the active damper, and the filter filters out the low-frequency harmonics so that the active damper only responds to problems within the resonant frequency band;
[0009] Step S4: The actual output impedance characteristic of the active damper is affected by (1+T A (s)) / T A (s) and the influence of the controller control effect, the active damper resonance suppression is difficult to achieve the ideal effect. The active damper using the output impedance reshaping strategy is used to convert the harmonic current reference i href_αβ With the sampling current i B_αβ Before adding, it is sent to an impedance reshaping stage to define the output impedance Z of the active damper BL (s) is the original port impedance Z o (s) and the virtual impedance Z VR (s) are connected in parallel, and the output impedance of the active damper is regarded as a resonance suppression whole, with the goal of making the output impedance of the active damper present a pure resistance characteristic externally;
[0010] Step S5: Propose an output impedance phase compensation strategy, use a non-ideal GI to replace the delay link, so that the active damper can maintain good pure resistance characteristics within the resonant frequency range at a switching frequency of 20kHz, and solve the contradiction between the switching frequency of the active damper and the output impedance resistance accuracy;
[0011] Step S6: Introduce the SOM neural network in the virtual impedance link, take the voltage resonance component of the common point as input, and the power of the active damper determines the screening limit. The screened resonance components with high voltage harmonic content are clustered according to the screening limit. After the optimization learning of SOM, the neighborhood neurons and the winning neurons are updated in real time towards the target with higher harmonic content, and the virtual resistance value is selected according to the objective function.
[0012] Compared with the prior art, the present invention has the following beneficial effects:
[0013] The active damper is only used to absorb the high-order harmonics of the system, so the resistor it simulates only needs to present low impedance within the target resonant frequency band, and there is no need to respond to low-order harmonics. The SOM neural network algorithm is introduced in the virtual resistance link to continuously update the weights of the input common point voltage resonant components. The higher the harmonic content, the greater the weight, and gradually achieve accurate mapping of the harmonic content of the input neurons, thereby outputting the common point voltage harmonic spectrum with a high harmonic content. According to the common point voltage spectrum with a high harmonic content, the input layer input common point voltage resonant components are clustered, and the winning neuron is judged by the winning neuron selection equation. The power of the active damper determines the limit of the screening resonant content, so that the resonant components with common point harmonic content higher than the limit are clustered, so that the neighboring neurons and the winning neurons are clustered towards the target of high harmonic content. The virtual resistor R v The value of is adaptively adjusted according to the harmonic component with the highest harmonic content. The proposed active damper impedance adaptive method based on SOM neural network can make the output impedance of the active damper present pure resistance characteristics at a lower switching frequency. The active damper has a good resonance suppression effect in a wide frequency band, reducing switching losses while increasing equipment power. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] Figure 1 It is a control block diagram of active damper using impedance adaptive method based on SOM neural network;
[0015] Figure 2 It is the equivalent control block diagram of capacitor voltage full feedback;
[0016] Figure 3 It is the mathematical model of active damper with total output impedance reshaping strategy;
[0017] Figure 4 is the equivalent circuit of the active damper port after impedance reshaping;
[0018] Figure 5 It is the Bode diagram of the digital delay link and the first-order inertia link;
[0019] Figure 6 This is the Z before and after impedance reshaping and phase compensation. BL / R tr Bird Chart;
[0020] Figure 7 It is the experimental waveform of inverter and active damper under strong power grid;
[0021] Figure 8 YesL g =5mH inverter experimental waveform and harmonic analysis;
[0022] Fig. 9 YesL g=5mH after adding active damper, the experimental waveform and harmonic analysis of the inverter;
[0023] Fig.10 YesL g =5mH active damper experimental waveform and harmonic analysis;
[0024] Fig.11 YesL g =5mH, inverter experimental waveform and harmonic analysis after active damper phase compensation;
[0025] Fig.12 YesL g =5mH, the active damper experimental waveform and harmonic analysis after phase compensation;
[0026] Fig.13 YesL g =9mH inverter experimental waveform and harmonic analysis after active damper phase compensation;
[0027] Fig.14 YesL g =9mH, experimental waveform and harmonic analysis after phase compensation of active damper;
[0028] Fig.15 It is the flow chart of the SOM algorithm of the neuron population optimization strategy;
[0029] Fig.16 This is the SOM network structure diagram. DETAILED DESCRIPTION
[0030] The technical solution of the present invention is described in detail below in conjunction with the principle block diagram and specific embodiments. It should be understood that the specific features of the embodiments of the present application are a detailed description of the technical solution of the present application, rather than a limitation of the technical solution of the present application. In the absence of conflict, the embodiments of the present application and the technical features in the embodiments may be combined with each other.
[0031] In a first aspect, the present invention provides an active damper impedance self-adaptation method based on a SOM neural network, comprising:
[0032] The neuron nodes in different areas of the competition layer of the SOM neural network have different divisions of labor and functions. Adjacent neurons can be activated by the common point voltage spectrum with similar distribution and then respond. If the size of a common point voltage resonance component in some neurons is different from the response of other neurons, then this output common point voltage resonance component can represent the category of the input common point voltage resonance component. Neurons adjacent to the activated neurons will be affected differently, and this influence is from far to near. If the neuron that is activated and sends a signal is taken as the center point, the neurons near the neuron react to the excited state through interaction. On the contrary, the neurons far away from the neuron will show an inhibitory state. If the distance is too far, the neuron will show a weak excitation state. The SOM neural network can adaptively determine the distribution of the common point voltage resonance component based on the different reactions of the nodes. Therefore, the SOM neural network relies on the different reactions of neurons to the common point voltage resonance component to realize the classification function. In addition, it can also make the common point voltage spectrum of the neurons that are stimulated by the same kind of excitation gather in a certain area in the output layer, thereby realizing the clustering function of the common point voltage resonance component.
[0033] The active damper impedance adaptive method based on SOM neural network needs to analyze the active damper topology structure and its control method to obtain its topology structure, wherein the topology structure is composed of the power switch tube Q A1 ~Q A6 Connected in reverse parallel with the diode to form a three-phase bridge inverter circuit, V dc The three-phase bridge inverter circuit is connected in parallel to the common grid point through the LCL filter to absorb the harmonics in the grid-connected inverter system. L1 is the inverter side inductor, C is the filter capacitor, and L2 is the grid side inductor. The grid impedance Z g (s) is usually inductive, where the resistance part can absorb harmonics, which is beneficial to the stability of the system, while the inductance will cause current lag, resulting in system instability. This article considers the worst case in the analysis, that is, the grid impedance is purely inductive, and the grid impedance is represented by Z g =L g Indicates that L g is the grid inductance, v g is the grid voltage.
[0034] The virtual resistance control loop is used to generate virtual resistance. The active damper collects the grid voltage from the public point and then passes through the notch filter G NA (s) Filter out the fundamental frequency and the 5th and 7th harmonic components, and retain the high-order harmonic components that need to be damped v pcch_αβ After being sent to the virtual resistance link and filtering out low-frequency harmonics, the harmonic current benchmark of the active damper is:
[0035] i href_αβ =vpcch_αβ / R v
[0036] Among them, R v is the virtual resistance value required to simulate the active damper. In order to effectively damp the voltage oscillation at the common point, the virtual resistance value of the active damper must be reasonably selected. Generally speaking, the smaller the virtual resistance value, the better the damping effect, and the more conducive to improving the stability of the system. However, if a very small virtual resistance value is directly set to ensure the damping effect, the active damper will absorb a large harmonic current from the power grid, increasing power loss.
[0037] Furthermore, the active damper port equivalent circuit is obtained. The virtual impedance of the active damper is simulated by inner loop control. By establishing its impedance model, its port impedance characteristics can be obtained. In the αβ coordinate system, the mathematical models of the α axis and the β axis are the same and decoupled. Taking the α axis component as an example, a mathematical model is established, in which the delay link It is used to indicate the 1.5-beat delay caused by calculation delay and modulation delay, T s is the sampling period of the active damper, T s =1 / f s , f s is the sampling frequency, thus we get
[0038]
[0039] Among them, G x1 (s) and G x2 (s) is the equivalent transfer function of the current loop; G de (s) is the control delay function; L1 and L2 are the inverter side inductance and grid side inductance parameters respectively; C is the filter capacitor parameter, K c is the capacitor current feedback coefficient. G pr (s) is the PR controller transfer function, and its expression is:
[0040]
[0041] Among them, K pA is the proportional control coefficient, K rA is the resonance control coefficient; ω0=2πf0, f0=50Hz, is the fundamental frequency; ω iA is the resonance bandwidth. To ensure that the inverter fundamental frequency can be stably connected to the grid when it fluctuates within the range of ±10%, ω iA =π. The LCL filter will introduce a resonant peak into the system, so adding the capacitor current i c_abc The feedback link suppresses the resonance peak, where K cA is the capacitor current feedback coefficient, v MA_abc is the modulation voltage.
[0042] The loop gain is:
[0043]
[0044] The expression of the active damper port current under the α axis is:
[0045]
[0046] Among them, v pcc (s) is the common point voltage, Z o (s) is the original port impedance of the active damper under the α axis, i ads (s) is the equivalent fundamental current source under the α axis, Z VR (s) is the impedance corresponding to the virtual resistance under the α-axis, and their expressions are:
[0047]
[0048] Among them, T A (s) is the loop gain, G NA (s) is the transfer function of the filter, It is the fundamental current reference.
[0049] According to the combined Norton theorem, the fundamental current source i ads (s) The harmonics absorbed by the active damper from the grid are fed back to the grid in the form of grid-connected current. In the port equivalent circuit of the active damper, Z o (s) corresponds to the original port impedance of the grid-connected part of the active damper, Z VR (s) corresponds to the virtual impedance introduced by the virtual resistance part.
[0050] Furthermore, the expression of the grid-connected point voltage of the port equivalent circuit of the active damping is:
[0051]
[0052] Among them, Z″ o (s) is the inverter equivalent output impedance Z eq (s) The original port impedance Z of the grid-connected part of the active damper o (s) and the virtual impedance Z introduced by the virtual resistance part of the active damper VR (s) and grid impedance Z g (s) parallel impedance value, Z o ″(s)=Z eq (s) / / Z o (s) / / Z VR (s) / / Z g (s), i s(s) is the inverter equivalent output DC source, i ads (s) is the fundamental current source, in a strong power grid environment, i s (s), i ads (s) contains only the fundamental component, v g (s) is the grid voltage, Z g (s) is the grid impedance value. Under weak grid conditions, v g (s) contains background harmonics, and the common point also contains low-order harmonics mainly composed of 3rd, 5th, and 7th harmonics. According to the expression of the common point terminal voltage, to eliminate the 3rd, 5th, and 7th harmonic components of the common point, Z″ must be satisfied. o (s)<<Z g (s), which requires the virtual resistance of the active damper to be as small as possible. However, the grid impedance varies in a wide range under weak grid conditions. g When it is close to zero, the virtual resistance value of the active damper cannot be infinitely reduced, and Z g (s) decreases with decreasing frequency. The harmonic suppression effect of the active damper at low frequencies is poor. When low-frequency harmonics flow through the active damper, additional energy loss will be generated, resulting in power waste.
[0053] According to the active damper, it is only used to absorb the high-order harmonics of the system. Therefore, the simulated resistor only needs to present low impedance in the target resonant frequency band, and there is no need to respond to low-order harmonics. If you want to suppress low-order harmonics at the grid connection point and improve the quality of grid-connected current, you can connect a harmonic compensation device such as LPF in series on the grid side. In the virtual resistance loop, add a filter G NA (s) to filter out the fundamental frequency and the 5th and 7th harmonics. Filter G NA The expression of (s) is:
[0054]
[0055] Among them, ω0 is the bandwidth at the fundamental frequency, ω0=2πf0, f0 is the fundamental frequency, f0=50Hz, h is the harmonic order corresponding to the characteristic frequency of the notch filter, and h=1, 5, 7.
[0056] According to the active damper, the resonance is damped by simulating the resistance characteristic in the target resonance frequency band. The actual output impedance characteristic of the active damper is affected by (1+T A (s)) / T A (s) and the effect of the controller control effect, which is not always equal to R in the resonant frequency band v, at this time, the resonance suppression of the active damper is difficult to achieve the ideal effect. When the phase of the virtual impedance of the active damper is shifted by 0°, the output impedance can be expressed as the real part plus the imaginary part. The imaginary part of the impedance may affect the actual resonance suppression effect of the active damper. Existing studies have pointed out that the active damper should present a pure resistance characteristic within the target resonance frequency band, but there is a lack of principle analysis of the influence of the imaginary part of the impedance on the harmonic suppression effect of the active damper.
[0057] Furthermore, based on the premise that the total impedance of the system should have a positive real part as the premise of system stability, the influence of the imaginary part of impedance on the resonance suppression effect is analyzed in principle, and the potential resonance risk that may be introduced by the imaginary part of impedance is explained from the perspective of system LC resonance;
[0058] According to the above assumptions, before adding the active damper, the total output impedance of the inverter is Z o (s), written as real part plus imaginary part:
[0059] Z inv (j2πf ust )=|Z inv |∠θ=R inv +jX inv
[0060] Among them, Z inv (j2πf ust ) is the total output impedance of the inverter is Z o (s) at the resonant frequency f ust The value at which the inv |Add phase angle ∠θ or real part R inv Add imaginary part X inv in the form of.
[0061] According to the above, in order to ensure that the inverter always maintains stability when the grid impedance changes, the real part of the total output impedance of the grid-connected system must be positive. Assume that before the active damper is put into use, the system itself already has stability problems. At the resonant frequency f ust At θ<-90°, that is, R inv <0,X inv <0. Assume that the active damper constructs a virtual impedance Z at the common point through the control algorithm VR (s) to suppress system resonance, Z VR (s) at the resonant frequency f ust The virtual impedance Z VR (s) can also be written in terms of real and imaginary parts:
[0062] Z VR (j2πf ust )=|Z VR |∠δ=R VR+jX VR
[0063] Among them, Z VR (j2πf ust ) is the virtual impedance Z VR (s) at the resonant frequency f ust The value at which the VR |Add phase angle ∠δ or real part R VR Add imaginary part X VR in the form of.
[0064] The total output impedance of the corrected system is:
[0065]
[0066] Further, we can simplify and write it in the form of real and imaginary parts as follows:
[0067]
[0068] Combined with the first aspect, further, the necessary and sufficient condition for the stability of the modified system is: Z eq (j2πf ust )’s phase θ′ satisfies -90°<θ′<90°, that is, the total output impedance Z eq (j2πf ust ) must have a positive real part.
[0069] According to the form of the real and imaginary parts of the total output impedance of the modified system, the denominator is in the form of a square sum, which is always greater than 0. If we want Z eq (j2πf ust ) has a positive real part requires:
[0070]
[0071] According to the resonant frequency f ust R inv <0,X inv <0, the imaginary part of impedance only exists in the last term In the case of virtual impedance, whether it is RC or RIS characteristics, That is, the existence of the imaginary part of the virtual impedance will produce a negative damping effect, which weakens the resonance suppression effect of the active damper. Negative damping characteristics, it is best to ensure X VR =0, that is, the virtual impedance of the active damper is a pure resistance.
[0072] According to the R VR As an independent variable, f(R VR ) is a quadratic function that opens downward. We want the quadratic function f(x) = ax2 +bx+c is a positive value, then b must be satisfied 2 -4ac>0,a=R inv , Z eq (j2πf ust ) has a positive real part is:
[0073]
[0074] The solution is: In order to ensure that R inv With X inv It is valid for any value. Should be less than its possible minimum value, satisfying That is, X VR =0.
[0075] According to the above, when x = -2a / b, f(x) can reach the maximum value, and a = R inv , Substitute x = -2a / b, when When Z eq (j2πf ust ) has a positive real part. VR ) has a maximum value of:
[0076]
[0077] According to the resonant frequency f ust At θ<-90°, R inv <0.X VR = 0, just take Regardless of the total impedance of the original system inverter R inv , X inv No matter how it changes, the total impedance Z of the corrected system can be guaranteed eq (j2πf ust )The real part is positive, which means that the system must be stable after adding the active damper.
[0078] According to the X VR The existence of makes the real part of the corrected total impedance have negative terms, X VR The larger the negative damping effect, the more obvious it will be, which will greatly weaken the resonance suppression effect of the active damper. VR Take the ideal value Too big X VR (resistance / capacitance is too strong), which may also cause Z eq (j2πf ust ) has a positive real part and the maximum value required is less than zero. The corrected total output impedance now has a negative real part, which cannot improve or even aggravates the stability problem of the system.
[0079] Secondly, according to the active damper impedance adaptive method based on the SOM neural network, to ensure its harmonic suppression effect, the virtual resistor should be kept as pure resistor as possible within the target damping frequency band, and the virtual resistor R v The smaller the value of , the better the damping effect and the better the system stability, but the power loss is large, so the virtual resistance is adaptively adjusted through the SOM neural network. At the same time, in view of the problem of excessively high switching frequency of the existing active damper, an output impedance phase compensation strategy is proposed. The compensated active damper still has a good harmonic suppression effect at a lower switching frequency, reduces switching losses, and improves the device power of the active damper. The specific process includes:
[0080] For each input, the SOM neural network adjusts some weights to make the weight closer to the input value. With this continuous learning and adjustment, the weights separate from each other and compete to achieve clustering.
[0081] The learning process of SOM is divided into 3 steps:
[0082] (1) Vector normalization
[0083] Different vectors have different lengths and directions. Vector normalization converts the vector into a unit vector with a constant direction and a length of 1. First, the current input pattern vector X in the self-organizing network and the inner star weight vector corresponding to each neuron in the competition layer are transformed into All are normalized to obtain and The vector normalization formula is:
[0084]
[0085] (2) Finding the winning neuron
[0086] When the network gets an input pattern vector When all neurons in the competition layer correspond to the inner star weight vector W j (j=1,2,...,m) are all the same as For similarity comparison, The most similar inner star weight vector is judged as the winning neuron of the competition, and its weight vector is recorded as The method of measuring similarity is to and Calculate the Euclidean distance. The Euclidean distance calculation formula is:
[0087]
[0088] (3) Network output and weight adjustment
[0089] The winning neuron outputs 1, and the rest outputs 0. That is:
[0090]
[0091] Among them, j (t+1) is the output of the winning neuron of the network at time t+1, To compete for the winning neuron, is the inner star weight vector corresponding to all neurons. * , that is, the inner star weight vector corresponding to the jth neuron is the winning neuron, and the network output is 1. When j≠j * , that is, the inner star weight vector corresponding to the j-th neuron is not the winning neuron, and the network output is 0.
[0092] Only the winning neuron has the right to adjust its weight vector The adjusted weight vector is:
[0093]
[0094] In the formula, is the weight vector of the neuron that wins the competition and is most similar to the weight vector of the inner star at time t+1, ΔW j* is the adjusted weight of the output neuron, is the learning rate, which decreases as learning progresses. It can be seen that when j≠j * When , the inner star weight vector corresponding to the jth neuron is not the winning neuron, and the weight of the corresponding neuron cannot be adjusted. In essence, the "winner" strongly inhibits them and does not allow them to be excited. The new vector obtained after the normalized weight vector is adjusted is no longer a unit vector, so the adjusted vector needs to be renormalized. After step (3) is completed, return to step (1) to continue training until the learning rate α decays to 0 or the specified value.
[0095] The specific competition mode on the output layer is: for the winning neuron t, the neurons in the surrounding area are excited to varying degrees, and the others are inhibited. When different samples are input from the outside, the positions of excited neurons caused by the input samples are different at first, but after self-organization, some specific neural groups will be formed, which represent the input samples and their characteristics respectively.
[0096] According to the active damper control method, the active damper adopts the output impedance reshaping strategy to convert the harmonic current reference i href-αβ With the sampling current i B-αβ Before adding, they are sent to an impedance reshaping stage.
[0097] According to the derivation of the active damper port equivalent electrical circuit, the active damper port equivalent impedance includes the original output impedance Z0(s) and the virtual impedance Z VR (s) There are two parts. When the traditional active damper provides virtual resistance for the system, only the virtual impedance Z is considered separately. VR (s) The suppression effect of harmonics. The existing virtual resistance compensation strategy is only based on the virtual impedance Z VR (s) is the optimization target, and the original output impedance Z0(s) is ignored for its harmonic suppression effect. Reasonable use of the original output impedance Z0(s) can make the active damper have a stronger harmonic suppression effect and improve the utilization efficiency of the active damper equipment. VR (s) Parallel form is the active damper output impedance Z BL (s), the harmonic suppression effect of the original output impedance Z0(s) is included in the optimization range, the output impedance of the active damper is regarded as a resonance suppression whole, an impedance reshaping method is proposed, and an active damper impedance reshaping strategy is proposed with the goal of making the output impedance of the active damper present a pure resistance characteristic to the outside.
[0098] After adding the impedance reshaping link, the original port impedance of the active damper is Z0(s), and the virtual impedance Z VR The expressions of (s) are:
[0099]
[0100] in:
[0101] According to the above, the original port impedance Z0(s) of the active damper is included in the optimization range to suppress harmonics. The output impedance Z BL (s) is the original port impedance Z0(s) and the virtual impedance Z VR The parallel connection of (s) is expressed as:
[0102]
[0103] According to the optimization goal of reshaping the output impedance of the active damper into a pure resistor, the expression of the impedance reshaping link is derived, and Z BL (s) is a pure resistor, namely Z BL (s) = R tr , we can solve it back to get G tr (s) is expressed as:
[0104]
[0105] According to the use of G tr(s) Reshape the output impedance, combined with Z BL (s), the expression of output impedance is:
[0106]
[0107] According to the Z BL The numerator and denominator of formula (s) both contain the common factor R V This shows that after impedance reshaping, the size of the total output impedance is independent of the virtual resistance value set in the virtual impedance loop. Therefore, the virtual resistance link can be omitted, and the actual total output impedance is determined by the parameter R in the impedance reshaping link. tr Decided to tr is defined as the impedance coefficient of the active damper after impedance reshaping. Considering the digital delay link There is a phase advance, which is difficult to achieve with actual physical methods. The influence of its phase is ignored and approximated to 1. Finally, the expression of the total output impedance reshaping link is:
[0108]
[0109] Among them, the notch filter and current PR controller are designed only for the fundamental and low-order harmonics, and have little effect on other frequency bands. Therefore, it can be considered that at the high-order harmonics damped by the active damper, there is approximately G pr (s)≈K p , G pr (s) is the transfer function of the PR controller.
[0110] In the existing methods for compensating the resistance accuracy of active dampers, the delay link G is often ignored. de (s), approximate it to 1, and consider the total output impedance of the delay link to reshape the link:
[0111]
[0112] It is found that when the switching frequency is 20kHz, if the delay link in the impedance reshaping link is still ignored, the virtual resistor will no longer present a pure resistance characteristic, weakening the harmonic suppression effect of the active damper. Therefore, the impedance reshaping link is re-optimized to compensate the phase of the output impedance: using the first-order inertia link G FOI (s)=1 / (1.5T s s+1)(T s The sampling period of the active damper) approximates the delay link, and the first-order inertia link restores the phase characteristics of the digital delay link to a certain extent, which can effectively compensate for the influence of the digital delay link on the phase of the total output impedance. FOI (s) Replace G de(s), after phase compensation, the expression of impedance reshaping link is:
[0113]
[0114] The above research can further improve the accuracy of the output impedance of the impedance reshaping strategy. After using the first-order inertia link to compensate the output impedance phase, the expression of the impedance reshaping link contains differential and second-order differential terms. The differential term is easily interfered by the system noise, and its digital realization is more difficult. Using non-ideal GI to replace the differential term can effectively reduce the noise, which is conducive to the realization of the differential term in digital control. The expression of non-ideal GI is:
[0115]
[0116] Among them, the cut-off frequency is ω c =8000rad / s. Select the angular frequency ω at the maximum gain according to the prototype parameters * =πf s , and use the FOH method to get GI * The discretization result of (s) is expressed as:
[0117]
[0118] Among them, GI * (z) is the expression of non-ideal GI * (s) Discretized expression.
[0119] It is found that within the target resonant frequency band of 1000-2000Hz, non-ideal GI can effectively replace the differential term and restore the differential frequency characteristics to a large extent. Using non-ideal GI to replace the differential term can greatly reduce the interference of noise, and this method is relatively simple to implement in a digital system.
[0120] The specific algorithm steps of the SOM network are as follows:
[0121] (1) Set input variables and weight vectors
[0122] The input variable is the common point voltage resonant component: K i (n)=K1(n),K2(n),...,K N (n);
[0123] Among them, K i (n) is the input variable, n is the number of learning times, N is the total number of learning times, and it is determined whether the number of learning times n exceeds N, otherwise n=n+1. At this point, all learning samples are provided to the neural network.
[0124] The weight vector is the common point voltage harmonic spectrum with a total harmonic distortion not exceeding 0.6%: Wij (n) = x 1j (n),x 2j (n),...,x Nj (n)
[0125] (2) Initialization: The connection weight vector W of the SOM network ij Assign random values in the interval [0,1] to determine the learning rate η(n) and the neighborhood N t The initial value η(0) (0<η(0)<1) and N t (0)
[0126] (3) For the input vector K i and weight vector W ij Perform unitization, that is:
[0127]
[0128] in, is the normalized input vector, K i is the input vector, is the normalized weight vector, W ij is the weight vector.
[0129] (4) Given any input sample K i , and calculate K i With W ij Euclidean distance between weight vectors:
[0130]
[0131] (5) Find the minimum distance d t , determine the winning neuron t:
[0132] d t =min[d j ]
[0133] (6) Adjust the connection weights to the output layer neighborhood N t The weight vectors between all neurons in (n) and the neurons in the input layer are modified:
[0134] W ij =(n+1)=W ij (n)+η(n)·[K i -W ij (n)]
[0135] (7) Select another sample and repeat steps (4)-(6) until all samples are learned. Get the new learning rate η(n) and topological neighborhood N t (n):
[0136]
[0137] In the formula, n is the number of learning times, N is the total number of learning times, and it is determined whether the number of learning times n exceeds N, otherwise n = n + 1, and all learning samples are provided to the neural network. Through the steps of this algorithm, the winning neuron t in the output layer and the weight vector in the neighborhood are finally made close to the input vector to obtain the best output.
[0138] According to the above, the proposed active damper impedance adaptive method based on SOM neural network improves the original winning element selection criterion by improving the formula of K value, and the selection criterion is:
[0139]
[0140] The present invention provides an active damper impedance adaptive method based on a SOM neural network. In this method, the active damper is only used to absorb the high-order harmonics of the system. Therefore, the simulated resistor only needs to present a low impedance in the target resonant frequency band, and there is no need to respond to low-order harmonics. The SOM neural network is introduced at the virtual resistor. Through the SOM neural network, the input layer inputs the common point voltage resonance component for clustering, and the winning neuron is judged by the selection equation of the winning neuron. The power of the active damper determines the screening limit, so that the common point resonance component is clustered towards the higher harmonic content, so that the neighboring neurons and the winning neurons are clustered towards the target with higher harmonic content, so that the virtual resistor R v Adaptive adjustment is performed based on the resonant component with the highest harmonic content. The proposed active damper impedance adaptive method based on SOM neural network can make the output impedance of the active damper present pure resistance characteristics at a lower switching frequency. The active damper has a good resonance suppression effect in a wide frequency band, reducing switching losses while increasing equipment power.
[0141] The control block diagram of the active damper using impedance reshaping strategy is as follows: Figure 1 As shown, the topology consists of a power switch tube Q A1 ~Q A6 Connected in reverse parallel with the diode to form a three-phase bridge inverter circuit, V dc The three-phase bridge inverter circuit is connected in parallel to the common grid point through the LCL filter to absorb the harmonics in the grid-connected inverter system. Among them, L1 is the inverter side inductance, which is 0.6mH in the present invention; C is the filter capacitor, which is 8μF in the present invention; L2 is the grid side inductance, which is 0.4mH in the present invention. The grid impedance is Z g =L g Indicates that v g is the grid voltage.
[0142] The control of the active damper is carried out in the stationary coordinate system α-β coordinate system. The virtual resistance control loop is used to generate the virtual resistance. The active damper collects the grid voltage v from the common point. pcc Then pass through the notch filter G NA (s) Filter out the fundamental frequency and the 5th and 7th harmonic components, filter G NA The expression of (s) is:
[0143]
[0144] Among them, ω0=2πf0, f0=50Hz, h is the harmonic order corresponding to the characteristic frequency of the notch filter, and h=5,7.
[0145] Retain the high-order harmonic components v that need to be damped pcch_αβ Then it is sent to the virtual resistance link, and the resonant components at the specific frequency of the common point are clustered through the SOM neural network. The active damper power P determines the threshold of the clustering amount, and the weight is determined according to the harmonic content. The winning neuron is judged by the winning neuron selection equation, so that it can be used to judge the common point voltage resonance and the active damper power loss, thereby determining the virtual resistance R v The value of harmonic current reference i href_αβ With the sampling current i B_αβ Before adding, it is first sent to the impedance reshaping link G tr (s) Get the harmonic current reference i of the source damper href_αβ =v pcch_αβ / R v , R v is the virtual resistance value determined by the SOM neural network. The inverter control strategy is carried out in the stationary coordinate system α-β coordinate system. The grid voltage at the PCC point is collected by a phase-locked loop, and the grid-connected current i B_αβ and harmonic current reference i href_αβ and then add it to the fundamental current reference The difference between them is used as the grid-connected current reference and sent to the PR controller. Its expression is:
[0146]
[0147] Among them, K pA is the proportional control coefficient, which is 8 in the present invention; K rA is the resonance control coefficient, which is taken as 12000 in the present invention; ω0=2πf0, f0=50Hz, which is the fundamental frequency; ω iA is the resonance bandwidth. To ensure that the inverter fundamental frequency can be stably connected to the grid when it fluctuates within the range of ±10%, ω iA =π. The LCL filter will introduce a resonant peak into the system, so adding the capacitor current i c_abcThe feedback link suppresses the resonance peak, where K cA is the capacitor current feedback coefficient, which is 25 in the present invention. The output signal of the PR modulator is subtracted from the filter capacitor current feedback signal to obtain the modulation voltage v MA_abc The modulated voltage is sent to the PWM modulator to generate a drive signal that controls the switch tube in the inverter.
[0148] The active damper damps the system resonance by simulating resistance in the resonant frequency band through the control strategy. Its port equivalent circuit is as follows: Figure 2 The port equivalent impedance of the active damper includes the original port impedance Z0 and the virtual impedance Z VR Therefore, there are two options for the impedance used by the active damper to damp resonance. One is to directly use the original port impedance Z0 to suppress resonance. Since the original port impedance is generated by the active damper grid-connected loop, its impedance characteristics will affect the grid-connected stability of the active damper. The only way to improve its suppression effect on resonance is to optimize the resonance parameters of the proportional resonant controller. The more widely used is to construct a virtual resistor Z by introducing a virtual resistor control loop. VR Damped resonance, the biggest advantage of using virtual resistance to damp resonance is that the virtual resistance control loop will not have any impact on the open-loop transfer function of the active damper. Additional compensation links can be designed to improve the accuracy of the virtual resistance without affecting the stability of the current loop of the active damper itself.
[0149] The mathematical model of the active damper using the output impedance reshaping strategy is as follows: Figure 3 After adding the impedance reshaping link, the original port impedance of the active damper is Z o (s), and the virtual impedance Z VR The expressions of (s) are:
[0150]
[0151] The original port impedance Z of the active damper is o (s) The harmonic suppression effect is included in the optimization range, and the output impedance Z of the active damper BL (s) is the port impedance Z o (s) and the virtual impedance Z VR The parallel connection of (s) is expressed as:
[0152]
[0153] Taking the output impedance of the active damper as a pure resistor as the optimization goal, the expression of the impedance reshaping link is derived, and Z BL (s) is a pure resistor, namely Z BL (s) = R tr , we can solve it back to get G tr(s) is expressed as:
[0154]
[0155] Utilize formula G tr (s) After reshaping the output impedance, the equivalent port circuit of the active damper is as follows: Figure 4 shown.
[0156] Figure 5 The Bode diagram of the non-ideal GI and its discretization results is shown. The non-ideal GI can effectively replace the differential term and restore the differential frequency characteristics to a large extent. Using non-ideal GI to replace the differential term can greatly reduce the interference of noise. This method is relatively simple to implement in digital systems.
[0157] Using the parameters of the active damper prototype, the impedance before and after impedance reshaping and after output impedance phase compensation were respectively made. BL / R tr Bird Figure 6 .
[0158] The figure compares the phase characteristics of the output impedance of the active damper when different control methods are used. Due to the addition of the notch filter, the fundamental wave and the 5th and 7th harmonic currents flowing into the active damper can be effectively suppressed, which reduces the power loss of the active damper to a certain extent. The total output impedance of the active damper should have a pure resistance characteristic, that is, the phase of the output impedance should be as close to 0° as possible. When the output impedance reshaping link is not added, the total output impedance phase characteristic of the active damper is shown in the red curve in the figure. Its phase begins to shift at the medium and low frequencies, and its output impedance shows a resistance-inductance characteristic in the resonant frequency band. Specifically, at 1000Hz, the output impedance phase is 61.8°. As the resonant frequency increases, the phase shift increases. At 2000Hz, the output impedance phase is even higher than 90°, which means that the output impedance of the active damper has a negative real part at this time, which cannot damp the resonance and even reduces the system stability margin.
[0159] After adding the impedance reshaping link, the phase shift of the output impedance is improved. As shown by the blue curve in the figure, at a frequency of 1000Hz, the phase is 13.9°. Compared with before impedance reshaping, the phase accuracy is improved by about 47°. However, at a switching frequency of 20kHz, the phase impact of the digital delay link on the output impedance gradually increases with the increase of frequency. At 2000Hz, the phase shift is 56.6°. At this time, the output impedance characteristic of the active damper is a resistance-inductance characteristic. At this time, the resonance damping effect of the active damper is not good, and there may be a potential resonance risk.
[0160] In order to make the output impedance present a more accurate pure resistance characteristic within the target resonant frequency band and improve the resonance suppression effect of the active damper, a compensation strategy for the output impedance phase is proposed, and the output impedance reshaping link is optimized. The compensation effect is as follows: Figure 6 As shown in the yellow curve, its phase characteristic is closer to 0° in the resonant frequency band of 1000-2000Hz, that is, the total output impedance of the active damper can always present good resistance characteristics to the outside. Specifically, the phase is 2.6° at 1000Hz, which is about 59° lower than the phase without adding the impedance reshaping link. At 2000Hz, the phase shift is only 6.9°, which is about 50° lower than before the phase compensation, and the compensation effect is significant.
[0161] The embodiment of the present invention is verified by experiments, and the parameters of the active damper are set as shown in Table 1.
[0162] Table 1 Grid-connected inverter parameters
[0163]
[0164] Figure 7 The grid-connected current waveforms of the inverter and active damper and the voltage waveform of the common point are given when the grid-side inductance is strong under the power grid.
[0165] When the grid strength decreases, if the grid impedance L g If the phase is lower than -90° at the intercept frequency where the amplitude intercepts with the inverter output impedance, it will cause inverter stability problems. Figure 8 Given L g =5mH (corresponding to SCR = 4.6) when the inverter grid-connected current and common point voltage experimental waveform and FFT analysis results of the grid-connected current. It can be seen from the figure that when L g =5mH, the output current and grid-connected voltage waveform of the inverter, which was originally able to operate stably, are seriously distorted under the influence of grid impedance and cannot meet the grid-connected stability requirements of the inverter. The results of FTT analysis of the grid-connected current at this time show that the THD value of the grid-connected current is 54.36%, the inverter cannot operate stably, and the harmonics are mostly concentrated in the frequency band of 1200Hz~1400Hz.
[0166] In order to solve the resonance problem in the system, the proposed active damper based on the impedance adaptive method of SOM neural network is connected in parallel at the common grid point. At this time, the inverter, active damper grid current and common point voltage, as well as the results of FFT analysis of grid current are as follows: Fig. 9 , Fig.10 shown.
[0167] contrast Fig. 9 and 10After adding an active damper using an impedance adaptive method based on a SOM neural network to the system, the quality of the inverter grid-connected waveform has been significantly improved, and the inverter grid-connected current THD value has dropped to 13.64%. Compared with before the active damper was added, the harmonic content has dropped by about 40%. This shows that the active damper using an impedance adaptive method based on a SOM neural network has suppressed the resonance phenomenon in the system to a great extent, significantly reducing the harmonic content of the system. However, at this time, the THD content of the grid-connected current is still high and does not meet the grid-connected requirement of THD < 8%. This is because the output impedance cannot fully guarantee the pure resistance characteristics within the resonant frequency band, and the harmonic suppression effect is weak, so the impedance reshaping strategy cannot achieve the ideal effect. Fig.10 It can be seen that at this time, the stability of the active damper is affected by the inverter, and the grid-connected current will also contain certain harmonic components. At this time, THD = 11.37%.
[0168] In order to solve the above problems, the resonance damping effect of the active damper is improved, the delay link is approximated by the first-order inertia link, the output impedance phase is compensated, and the impedance reshaping link is re-optimized. The control algorithm of the active damper is modified, and the active damper after the output impedance phase compensation is reintegrated into the system. At this time, the inverter, active damper grid-connected current and common point voltage, and the results of the grid-connected current FFT analysis are as follows: Fig.11 , Fig.12 shown.
[0169] from Fig.11 , 12 It is obvious that the grid-connected current waveform quality of the inverter and active damper is good at this time, and the harmonic content is significantly reduced compared to before phase compensation. Specifically, the THD value of the inverter grid-connected current drops to 2.55%, meeting the grid-connected requirements. Compared with before the active damper was added, the THD dropped by about 52%, and the harmonic content was reduced by 11% compared to before phase compensation. The grid-connected current harmonic content of the active damper is THD = 2.78%, which is 8.5% lower than the THD value before phase compensation. Through before and after comparison, it is proved that the active damper output impedance phase compensation method can play a role in phase compensation, making the output impedance of the active damper closer to the pure resistance characteristic in the target resonant frequency band, greatly improving the resonance suppression effect of the active damper.
[0170] In order to verify that the proposed active damper control strategy has good harmonic suppression capability in a wider target frequency band, the grid impedance is increased to L g =9mH to simulate the environment of SCR=2.5. At this time, the inverter, active damper grid-connected current and common point voltage, as well as the results of FFT analysis of grid-connected current are as follows: Fig.13 , Fig.14 shown.
[0171] from Fig.13 , 14 It can be seen that even in the environment of SCR<3, the inverter and active damper still have a good grid-connected current waveform. At this time, the THD of the grid-connected current of the inverter and active damper are 1.89% and 1.96% respectively. Slightly lower than L g = THD when 5mH, this is because L g When L increases g The interface frequency with the system impedance shifts to the low frequency region. g =9mH corresponds to the cutoff frequency, the phase of the inverter gradually increases, compared with L g =5mH corresponds to the phase at the cutoff frequency. The phase margin of the inverter itself is improved and the severity of the resonance is reduced.
[0172] Through before-and-after comparison experiments, the suppression effect of the active damper using the impedance adaptive method based on the SOM neural network on system harmonics and the effectiveness of the proposed output impedance phase compensation method are verified. The active damper using the proposed impedance adaptive method based on the SOM neural network can effectively suppress system resonance within a wide frequency band and improve the stability of the grid-connected inverter system under weak power grids.
[0173] Fig.15 The flow chart of the SOM algorithm with a population optimization strategy is given. SOM is an important method in unsupervised learning methods, which can be used for clustering, high-dimensional visualization, data compression, feature extraction and many other purposes. Due to its intuitive structure and clear mapping relationship, it is widely used in task planning. A SOM algorithm with a more efficient learning strategy for population optimization is proposed to improve the efficiency and speed of the system in task scenarios of different scales.
[0174] Fig.16 The SOM network structure diagram is given. In the input layer, each output neuron corresponds to a task, and the weight of the input neuron is the three-dimensional coordinate value of the task. In the output layer, each output neuron corresponds to a ring neuron structure, which is composed of a certain number of output neurons, and each output neuron represents a possible path point. After the algorithm learning is completed, only the real path points will remain to form a new ring neuron structure to complete all the assigned tasks.
[0175] The above is only a preferred implementation of the present invention. It should be noted that, for those skilled in the art, several improvements and changes can be made without departing from the technical principles of the present invention. These improvements and changes should also be regarded as the protection scope of the present invention.
Claims
1. An active damper impedance adaptive method based on SOM neural network, characterized in that: include: Step S1: the active damper damps the voltage resonance of the common point by constructing a virtual impedance, and the virtual impedance presents a resistance characteristic to the outside; Step S2: Establish a SOM neural network consisting of an input layer and a competition layer, and the input layer receives the resonant component from the common point; The competitive layer uses the methods of analysis, comparison and discovery to select the resonant components greater than the screening limit according to the selection equation, and obtains a resonant frequency distribution with a high content; Step S3: the active damper adds a filter in the virtual resistance loop to eliminate the interference of low-frequency harmonics on the function of the active damper, and the filter filters out the low-frequency harmonics so that the active damper only responds to problems within the resonant frequency band; Step S4: The actual output impedance characteristic of the active damper is affected by (1+T A (s)) / T A (s) and the influence of the controller control effect, the active damper resonance suppression is difficult to achieve the ideal effect; the active damper using the output impedance reshaping strategy is used to convert the harmonic current reference i href_αβ With the sampling current i B_αβ Before adding, it is sent to an impedance reshaping stage to define the output impedance Z of the active damper BL (s) is the original port impedance Z o (s) and the virtual impedance Z VR (s) are connected in parallel, and the output impedance of the active damper is regarded as a resonance suppression whole, with the goal of making the output impedance of the active damper present a pure resistance characteristic externally; Step S5: Propose an output impedance phase compensation strategy, use a non-ideal GI to replace the delay link, so that the active damper can maintain good pure resistance characteristics within the resonant frequency range at a switching frequency of 20kHz, and solve the contradiction between the switching frequency of the active damper and the output impedance resistance accuracy; Step S6: Introduce the SOM neural network in the virtual impedance link, take the voltage resonance component of the common point as input, and the power of the active damper determines the screening limit. The screened resonance components with high voltage harmonic content are clustered according to the screening limit. After the optimization learning of SOM, the neighborhood neurons and the winning neurons are updated in real time towards the target with higher harmonic content, and the virtual resistance value is selected according to the objective function.
2. The active damper impedance adaptive method based on SOM neural network according to claim 1 is characterized in that: Analyze the active damper topology, which consists of the power switch tube Q A1 ~Q A6 Connected in reverse parallel with the diode to form a three-phase bridge inverter circuit, V dc is the DC side voltage, and the three-phase bridge inverter circuit is connected in parallel to the public grid point through the LCL filter to absorb the harmonics in the grid-connected inverter system; where L1 is the inverter side inductor, C is the filter capacitor, and L2 is the grid side inductor; the grid impedance Z g (s) is usually inductive, where the resistance part can absorb harmonics, which is beneficial to the stability of the system, while the inductance will cause current lag, resulting in system instability. This article considers the worst case in the analysis, that is, the grid impedance is purely inductive, and the grid impedance is represented by Z g =L g Indicates that L g is the grid inductance, v g is the grid voltage; The virtual resistance control loop is used to generate virtual resistance. The active damper collects the grid voltage from the public point and then passes through the notch filter G NA (s) Filter out the fundamental frequency and the 5th and 7th harmonic components, and retain the high-order harmonic components that need to be damped v pcch_αβ After being sent to the virtual resistance link and filtering out low-frequency harmonics, the harmonic current benchmark of the active damper is: i href_αβ =v pcch_αβ / R v Among them, R v The virtual resistance value required to simulate the active damper.
3. The active damper impedance adaptive method based on SOM neural network according to claim 2 is characterized in that: Before adding the active damper in step S1 above, the total output impedance of the inverter is Z o (s), written as real part plus imaginary part: WITH inv (j2πf ust )=|From inv |∠θ=R inv +jX inv Among them, Z inv (j2πf ust ) is the total output impedance of the inverter is Z o (s) at the resonant frequency f ust and convert it into magnitude |Z inv |Add phase angle ∠θ or real part R inv Add imaginary part X inv form; To ensure that the inverter remains stable when the grid impedance changes, the real part of the total output impedance of the grid-connected system must be positive; assuming that the system itself has stability problems before the active damper is put into use, the resonance frequency f ust At θ<-90°, that is, R inv <0,X inv <0; Assume that the active damper constructs a virtual impedance Z at the common point through the control algorithm VR (s) to suppress system resonance, Z VR (s) at the resonant frequency f ust The virtual impedance Z VR (s) is also written in the form of real and imaginary parts, and its expression is: WITH VR (j2πf ust )=|From VR |∠δ=R VR +jX VR Among them, Z VR (j2πf ust ) is the virtual impedance Z VR (s) at the resonant frequency f ust The value at which the VR |Add phase angle ∠δ or real part R VR Add imaginary part X VR form; The total output impedance of the corrected system is Z eq (j2πf ust ), whose expression is: Written in real and imaginary form, the expression is: The necessary and sufficient conditions for the stability of the corrected system are: eq (j2πf ust )’s phase θ′ satisfies -90°<θ′<90°, that is, the total output impedance Z eq (j2πf ust ) must have a positive real part; Formula Z eq (j2πf ust ) The denominator is in the form of a sum of squares, which is always greater than 0. If you want Z eq (j2πf ust ) has a positive real part requires: At the resonant frequency f ust R inv <0,X inv <0, the imaginary part of impedance only exists in the last term In the case of virtual impedance, whether it is RC or RIS characteristics, That is, the existence of the imaginary part of the virtual impedance will produce a negative damping effect, which weakens the resonance suppression effect of the active damper. Negative damping characteristics, it is best to ensure X VR =0, that is, the virtual impedance of the active damper is a pure resistance.
4. The active damper impedance adaptive method based on SOM neural network according to claim 3 is characterized in that: The SOM neural network described in step 2 is divided into two layers, the upper layer is the competition layer, also called the output layer, and the lower layer is the input layer; the input layer of the input sample data has a total of h neurons, the number of neuron nodes in the input layer depends on the complex harmonic distribution, and the output nodes of the output layer are two-dimensional quadrilateral node matrices, which make appropriate responses by simulating the human cerebral cortex; The neurons in the competition layer are distributed in an m×n matrix, with a total of m×n neurons. There are two types of connection weights in the SOM neural network. One is the weight vector connecting the input layer and the competition layer, which reflects the connection weight of the neuron's response to external input; the other is the connection weight between neurons, and its size can control the strength of the connection between neurons; each neuron is connected to the surrounding neurons; when the neural network is in input mode, each neuron in the output layer will respond, and then adjust the voltage and grid-connected current spectrum at the common point of the weights to the direction of the winning neuron competition, thereby finally obtaining the resonant component with high harmonic content.
5. The active damper impedance adaptive method based on SOM neural network according to claim 4 is characterized in that: The active damper port equivalent circuit is obtained. The virtual impedance of the active damper is simulated by inner loop control. Its port impedance characteristics can be obtained by establishing its impedance model. In the αβ coordinate system, the mathematical models of the α axis and the β axis are the same and decoupled. Taking the α axis component as an example, a mathematical model is established, in which the delay link It is used to indicate the 1.5-beat delay caused by calculation delay and modulation delay, T s is the sampling period of the active damper, T s =1 / f s , f s is the sampling frequency, thus we get Among them, G x1 (s) and G x2 (s) is the equivalent transfer function of the current loop; G de (s) is the control delay function; L1 and L2 are the inverter side inductance and grid side inductance parameters respectively; C is the filter capacitor parameter, K c is the capacitor current feedback coefficient; G pr (s) is the PR controller transfer function, and its expression is: Among them, K pA is the proportional control coefficient, K rA is the resonance control coefficient; ω0=2πf0, f0=50Hz, is the fundamental frequency; ω iA is the resonance bandwidth. To ensure that the inverter fundamental frequency can be stably connected to the grid when it fluctuates within the range of ±10%, ω iA =π; the LCL filter will introduce a resonant peak into the system, so adding the capacitor current i c_abc The feedback link suppresses the resonance peak, where K cA is the capacitor current feedback coefficient, v MA_abc is the modulation voltage; The loop gain is: The expression of the active damper port current under the α axis is: Among them, v pcc (s) is the common point voltage, Z o (s) is the original port impedance of the active damper under the α axis, i ads (s) is the equivalent fundamental current source under the α axis, Z VR (s) is the impedance corresponding to the virtual resistance under the α-axis, and their expressions are: Among them, T A (s) is the loop gain, G NA (s) is the transfer function of the filter, It is the fundamental current reference.
6. The active damper impedance adaptive method based on SOM neural network according to claim 5, characterized in that: The port equivalent circuit of the active damping and the grid-connected point voltage are expressed as follows: Among them, Z o ″(s) is the inverter equivalent output impedance Z eq (s) The original port impedance Z of the grid-connected part of the active damper o (s) and the virtual impedance Z introduced by the virtual resistance part of the active damper VR (s) and grid impedance Z g (s) parallel impedance value, Z″ o (s) = Z eq (s) / / Z o (s) / / Z VR (s) / / Z g (s), i s (s) is the inverter equivalent output DC source, i ads (s) is the fundamental current source, in a strong power grid environment, i s (s), i ads (s) contains only the fundamental component, v g (s) is the grid voltage, Z g (s) is the grid impedance value. Under weak grid conditions, v g (s) contains background harmonics, and the common point also contains low-order harmonics mainly composed of 3rd, 5th, and 7th harmonics. According to the expression of the common point terminal voltage, to eliminate the 3rd, 5th, and 7th harmonic components of the common point, Z″ must be satisfied. o (s)<<Z g (s).
7. The active damper impedance self-adaptation method based on SOM neural network according to claim 6, characterized in that: In the virtual resistance loop, add a filter G NA (s) to filter out the fundamental frequency and the 5th and 7th harmonics; filter G NA The expression of (s) is: Among them, ω0 is the bandwidth at the fundamental frequency, ω0=2πf0, f0 is the fundamental frequency, f0=50Hz, h is the harmonic order corresponding to the characteristic frequency of the notch filter, and h=1, 5, 7.
Citation Information
Cited By
Three-phase inverter grid-connected oscillation suppression method and system
CN121036112A
A method and system for suppressing grid-connected oscillation in a three-phase inverter
CN121036112B
