Model-free predictive control method for parallel inverter based on neural network observer

Abstract

The full name of the application is: parallel inverter model-free predictive control method based on neural network observer. In order to solve the robust control problem of T-type three-level parallel grid-connected inverter system under filter inductance parameter mismatch, the application adopts reference value correction midpoint voltage control, neural network observer and simplified sequential model predictive control. The reference value correction midpoint voltage realizes the feedback control of the midpoint voltage independent of the filter inductance. The neural network observer estimates the uncertainty of the super-local model of the grid-side current and the zero-sequence circulating current. The simplified sequential model predictive control designs two cost functions of controlling the zero-sequence circulating current first and then controlling the grid-connected current, and simplifies the operation through the voltage vector grouping table, and finally outputs the optimal control signal of the inverter. The application makes the inverter under parameter mismatch to balance the midpoint voltage, fully suppress the zero-sequence circulating current, and output high-quality total grid-connected current, and improves the robustness of the grid-connected control system.

Description

Technical Field

[0001] This invention belongs to the field of power electronic converter control technology, specifically relating to a model-free predictive control method for parallel inverters based on a neural network observer. Background Technology

[0002] Currently, grid-connected inverters, serving as a bridge to convert direct current into high-quality alternating current, play a crucial role in distributed generation systems for new energy sources. T-type three-level inverters, with their lower total harmonic distortion (THD) and higher power output, are widely used in low-voltage applications. Furthermore, parallel grid-connected inverter systems can increase capacity and further reduce harmonic content.

[0003] In recent years, the robustness of power converter control algorithms has received considerable attention. For the filter stage of a grid-connected inverter system, if the inductor or capacitor experiences a mismatch between its actual and nominal values ​​due to aging or faults (i.e., parameter mismatch), the controller's performance will deteriorate. For a T-type three-level parallel inverter system, parameter mismatch not only increases the grid-side current distortion rate but also causes neutral-point voltage imbalance and an increase in zero-sequence circulating current. The purpose of robustness research is to design control algorithms that can still guarantee high-quality output characteristics under parameter mismatch.

[0004] Model-free predictive control (MBB) is a robust control method proposed in recent years. Its predecessor, model predictive control (MMC), has advantages such as simple design, fast dynamic response, and ease of implementing multi-objective control. However, it lacks robustness due to its reliance on an accurate mathematical model. MBB overcomes this shortcoming by improving robustness through methods such as designing hyperlocal models and estimation algorithms for their uncertainties.

[0005] Hyperlocal models are a simplified mathematical description of continuous nonlinear systems. They contain one or more uncertainties that need to be estimated using methods such as integral algebra, sliding mode observers, and neural network observers. Since these estimation methods do not depend on the system parameters but only on the sampled input-output information, control methods based on hyperlocal models will not be affected by parameter mismatch. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide a model-free predictive control method based on a neural network observer to achieve robust control of a T-type three-level parallel grid-connected inverter system when the filter inductor parameter mismatch occurs.

[0007] To achieve the above objectives, the present invention adopts the following technical solution.

[0008] Based on the topology diagram shown in Figure 1, the three-phase grid-side current of the two inverters is collected using voltage and current sensors. i 1. The voltage of the two capacitors on the DC bus side. u p and u n Zero-sequence circulation i z Defined as the sum of the three-phase grid-side currents of the first inverter. Neutral point voltage. u np Defined as u n reduce u p .

[0009] The overall system block diagram of this invention is shown in Figure 2. The controller consists of three parts: a midpoint voltage control based on reference value correction, a neural network observer, and a simplified sequential model predictive control.

[0010] The following is a neutral point voltage control based on a reference value correction. Let the corrected three-phase grid-side current reference value be... i 1arefM , i 1brefM , i 1crefM The corrected formula is: in I ref The magnitude of the grid-side current reference value. oh For the power grid frequency, k np The gain is controlled by the midpoint voltage.

[0011] Equations 1 to 3 do not involve the calculation of the filter inductance, thus achieving robust control of the midpoint voltage.

[0012] Below is the neural network observer. First, a hyperlocal model of the grid-side current is constructed: in F For uncertain terms, α=V dc / 2 L 1, u This is the voltage vector of the inverter. L 1 is the filter inductor. V dc This is the DC bus voltage.

[0013] Next, a hyperlocal model of the zero-order circulation is constructed: in f This is an uncertain term.

[0014] Then, the uncertain term in Equation 4 F The neural network observation formula for each sampling time is: Where the subscript E represents the observed value, W Represents the weight coefficient matrix. T s Represents the sampling time. f ( x )=1 / [+exp(- x ] is the activation function. k 1. t 1. k w1 This is a positive real number that needs to be set.

[0015] Finally, the uncertain term in Equation 5 f The neural network observation formula for each sampling time is: in, w Represents the weighting coefficient. k 2. t 2. k w2 This is a positive real number that needs to be set.

[0016] Below is a simplified sequential model predictive control. First, the uncertainty estimates are calculated by the neural network observer at each sampling time. F E ( k )and f E ( k The cost functions for grid-side current and zero-sequence circulating current are constructed as follows: Where || represents the vector 2-norm, i 1refM =[ i 1arefM , i 1brefM , i 1crefM ] T .

[0017] Secondly, in Formula Thirteen, u ( k There are 27 possible outcomes, and [1 1 1]u ( k All cases can be divided into seven groups based on the magnitude of their values. To reduce computation, one vector from each of the seven groups is selected and substituted. J z In the process, select those that make J z The three sets of vectors—the smallest, the second smallest, and the third smallest—are used to control the zero-sequence circulation suppression.

[0018] Finally, the three sets of vectors obtained are substituted into the cost function. J i In the middle, requesting an envoy J i The smallest switching vector is denoted as u opt ( k This switching vector will be converted into the on / off signal of the switching devices of the T-type three-level grid-connected inverter and sent in at the next sampling time.

[0019] Due to the application of the above technical solution, the present invention has the following characteristics: 1. This invention employs model predictive control technology, eliminating the need for PWM modulation. It can ensure the neutral point voltage balance and suppress zero-sequence circulating current in a T-type three-level parallel grid-connected inverter system, and improve the output quality of grid-side current. 2. This invention employs sequential predictive control technology, which, compared to traditional model predictive control, not only avoids the selection of weight factors but also reduces the number of loop operations, thus alleviating the burden on the processor. 3. This invention employs a midpoint voltage control based on reference value correction and a neural network observer, which can achieve robustness under filter inductor parameter mismatch. Attached Figure Description

[0020] Figure 1: Topology diagram of the T-type three-level grid-connected inverter in this invention;

[0021] Figure 2: Block diagram of model-free predictive control based on neural network observer in this invention;

[0022] Figure 3: Experimental waveform of three-phase total grid-connected current under parameter mismatch in this invention;

[0023] Figure 4: Experimental waveform of zero-sequence circulating current under parameter mismatch in this invention;

[0024] Figure 5: Experimental waveform of midpoint voltage under parameter mismatch in this invention. Detailed Implementation

[0025] The technical solution will be clearly and completely described below with reference to preferred embodiments and accompanying drawings. It should be understood that the preferred embodiments are merely illustrative of the invention and not intended to limit the scope of protection of the invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of the present invention.

[0026] This invention provides a robust control method for a T-type three-level parallel grid-connected inverter system under filter inductor parameter mismatch. By designing a reference-value-corrected midpoint voltage control and a neural network observer, this control method is unaffected by changes in the filter inductor value. The overall scheme can ensure the balance of the midpoint voltage, excellent total grid current quality, and sufficient suppression of zero-sequence circulating current even under parameter mismatch, thus improving the robustness of the parallel inverter system.

[0027] A schematic diagram of the control structure of one embodiment is shown in Figure 2. The main contents of this embodiment include the following steps:

[0028] Step S1: Collect the three-phase grid-side current of the two inverters using voltage and current sensors. i 1. The voltage of the two capacitors on the DC bus side. u p and u n Zero-sequence circulation i z Defined as the sum of the three-phase grid-side currents of the first inverter. Neutral point voltage. u np Defined as u n reduce u p Total grid-connected current i g Defined as the sum of the three-phase grid-side currents of the two inverters.

[0029] Step S2: Use the midpoint voltage u np Corrected three-phase grid-side current reference value i 1arefM , i 1brefM , i 1crefM The corrected formula is: in I ref The magnitude of the grid-side current reference value. oh For the power grid frequency, k np The gain is controlled by the midpoint voltage.

[0030] Step S3: Use delay element z -1 Obtain the three-phase grid-side current i 1 and zero sequence circulation i z The value at the previous sampling time.

[0031] Step S4: i 1( k )and i 1( k- 1) Substitute the values ​​into the neural network observer to obtain the estimated value of the uncertainty term. F E ( k The calculation formula is: Where the subscript E represents the observed value, W Represents the weight coefficient matrix. α=V dc / 2 L 1, u The voltage vector of the inverter. T s Represents the sampling time. f ( x )=1 / [+exp(- x ] is the activation function. k 1. t 1. k w1 This is a positive real number that needs to be set. L 1 is the filter inductor. V dc This is the DC bus voltage.

[0032] Step S5: ... i z ( k )and i z ( k- 1) Substitute the values ​​into the neural network observer to obtain the estimated value of the uncertainty term. f E ( k ). The calculation formula is: in, w Represents the weighting coefficient. k 2. t 2. k w2 This is a positive real number that needs to be set.

[0033] Step S6: Set [1 1 1] u according tou The values ​​obtained from the 27 cases are divided into 7 groups, as shown in Table 1.

[0034] Step S7: Substitute the 7 representative vectors of the 7 groups in Table 1 into the cost function. J z In the process, select those that make J z The three smallest, second smallest, and third smallest vectors. The cost function is:

[0035] Step S8: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full J z The three vectors are used to reconstruct several vectors according to Table 1, forming a set of candidate vectors.

[0036] Step S9: Substitute all the candidate vectors obtained in step S8 into the cost function. J i In the middle, requesting an envoy J i The smallest voltage vector is denoted as u opt ( k The cost function is as follows: Where || represents the vector 2-norm, i 1refM =[ i 1arefM , i 1brefM , i 1crefM ] T .

[0037] Table 1 shows the grouping of voltage vectors.

[0038] Step S10: Calculate the values ​​of the two inverters using the steps above. u opt ( k The signal is converted into the on / off signal of the switching devices of the two inverters and sent in at the next sampling time. Step S1 is executed again at the next sampling time.

[0039] Effect Experiment Example

[0040] To verify the effectiveness of the proposed model-free predictive control based on a neural network observer, a parallel grid-connected inverter system using the control scheme shown in Figure 2 is used as an example. The actual value of the filter inductance of the first inverter is 5mH, and the actual value of the filter inductance of the second inverter is 12mH. The control parameters of the proposed method are as follows: k np =0.1, k 1= k 2 = 10000 k w1 = k w2 =0.01, t 1= t 2=100. The grid-connected inverter and grid parameters are shown in Table 2.

[0041] Table 2 shows the parameters of the two grid-connected inverters and the power grid.

[0042] from Figure 3~Figure 5 As can be seen, under parameter mismatch conditions, the waveform quality of the three-phase grid-side current is very high, with a measured total harmonic distortion rate of 1.89%, which meets national standards. The peak-to-peak value of the zero-sequence circulating current is within 0.8A, demonstrating significant suppression. The maximum value of the neutral point voltage is only 2.2V, with very small fluctuations. The above experimental results demonstrate that this invention can ensure high-quality inverter performance, neutral point voltage balance, and zero-sequence circulating current suppression in a T-type three-level parallel grid-connected inverter system under parameter mismatch conditions.

[0043] The above embodiments are merely illustrative of the technical concept and features of the present invention, intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein. All equivalent changes or modifications made according to the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A model-free predictive control method for parallel inverters based on neural network observers, characterized in that, The method includes: Step S1: Collect the three-phase grid-side current of the two inverters using voltage and current sensors. i 1. The voltage of the two capacitors on the DC bus side. u p and u n Zero-sequence circulation i z Defined as the sum of the three-phase grid-side currents of the first inverter, the neutral point voltage u np Defined as u n reduce u p ; Step S2: Use the midpoint voltage u np Corrected three-phase grid-side current reference value i 1arefM , i 1brefM , i 1crefM The corrected formula is: in, I ref The magnitude of the grid-side current reference value. ω For the power grid frequency, k np Gain is controlled by the midpoint voltage; Step S3: Use delay element z -1 Obtain the three-phase grid-side current i 1 and zero-sequence circulation i z The value at the previous sampling time; Step S4: i 1( k )and i 1( k -1) Substitute this into the neural network observer to obtain the estimate of the uncertainty term. F E (k), the calculation formula is: Among them, subscript E Represents the observed value. W Represents the weight coefficient matrix. α = V dc / 2 L 1, u The voltage vector of the inverter. T s Represents the sampling time. φ (x)=1 / [+exp(- x ] is the activation function. k 1. τ 1. k w1 For the positive real number that needs to be set, L 1 is the filter inductor. V dc This is the DC bus voltage; Step S5: i z ( k )and i z ( k -1) Substitute this into the neural network observer to obtain the estimate of the uncertainty term. f E ( k ); The calculation formula is: in, w Represents the weighting coefficient. k 2. τ 2. k w2 The positive real number that needs to be set; Step S6: Place [1 1 1] u according to u The values ​​obtained from the 27 cases are divided into 7 groups; Step S7: Substitute the 7 representative vectors from the 7 groups into the cost function. J z In the process, select those that make J z The cost function is defined as follows: (The three vectors are: minimum, second minimum, and third minimum.) Step S8: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full J z The three vectors are restored to several vectors, forming a set of candidate vectors; Step S9: Substitute all the candidate vectors obtained in step S8 into the cost function. J i In the middle, requesting an envoy J i The smallest voltage vector is denoted as u opt ( k The cost function is: Where || represents the vector 2-norm, i 1refM =[ i 1arefM , i 1brefM , i 1crefM ] T ; Step S10: Calculate the values ​​of the two inverters using the steps above. u opt ( k The signal is converted into the on / off signal of the switching devices of the two inverters and sent in at the next sampling time. Step S1 is executed again at the next sampling time.

Images