A Neural Network Impedance Modeling Method Using Controller Information

Abstract

The present invention discloses a neural network impedance modeling method using controller information, the method is as follows: Step 1, obtain Z pn 、CV pn and CI pn ; Step 2, first stage training: input is frequency f, d-axis voltage V d , q-axis voltage V q , d-axis current I d , q-axis current I q , the output is the control information CV pp 、CV pn 、CV np 、CV nn , CI pp , CI pn , CI np , CI nn The amplitude and phase angle of the second stage training: the controller information CV trained in the first stage pn and CI pn The training results and frequency f, d-axis voltage V d , q-axis voltage V q , d-axis current I d , q-axis current I q Together they serve as the input of the second-stage neural network, and the output is the sequence impedance Z pp , Z pn , Z np , Z nn The present invention introduces control information into the impedance modeling process, which can improve the accuracy of neural network modeling.

Description

Technical Field

[0001] The invention relates to a neural network impedance modeling method, in particular to a neural network impedance modeling method using controller information. Background Art

[0002] The impedance-based analysis method is an important tool for studying the broadband stability problem of grid-converter interaction. Traditional impedance modeling requires obtaining all control structure, parameter and other information. Faced with black / gray models with unknown parameters, the actual impedance can only be obtained through frequency sweep measurement. Neural networks can simulate impedance models under multiple operating points through training. The accuracy of impedance modeling is an important prerequisite for stability analysis. However, the existing neural network-based converter impedance fitting method does not consider controller information and uses black / gray box models with low accuracy and large amount of data required.

[0003] When the manufacturer actually produces the equipment, the controller signal can be measured. By using the controller information, the correlation between the input and output data of the neural network can be increased, which can improve the accuracy of the neural network impedance modeling method. At present, there is no technical solution for how to use the controller information in the neural network impedance modeling. Summary of the invention

[0004] In view of the problem of low accuracy of the existing converter impedance fitting method based on neural network, the present invention provides a neural network impedance modeling method using controller information.

[0005] The objective of the present invention is achieved through the following technical solutions:

[0006] A neural network impedance modeling method using controller information includes the following steps:

[0007] Step 1: Control Information Modeling

[0008] Step 11: Obtain the sequence impedance Z of the voltage source converter pn

[0009] Sequence impedance Z pn The calculation formula is:

[0010]

[0011] In the formula, I p1 ,I n1 is the current obtained during the first group of frequency sweeps, V p1 、V n1 is the voltage obtained during the first group of frequency sweeps; I p2 ,I n2 is the current obtained during the second frequency sweep, V p2 、V n2 is the voltage obtained during the second group of frequency sweep; Zpp is the positive sequence voltage transfer function corresponding to the positive sequence current, Z pn is the positive sequence voltage transfer function corresponding to the negative sequence current, Z np is the negative sequence voltage transfer function corresponding to the positive sequence current, Z nn is the negative sequence voltage transfer function corresponding to the negative sequence current;

[0012] Step 12: Get controller information CV pn and CI pn

[0013] Control information CV pn and CI pn Defined as:

[0014]

[0015] In the formula, C p1 , C n1 is the control modulation wave obtained during the first group of frequency sweep, V p1 、V n1 is the voltage obtained during the first group of frequency sweeps; C p2 , C n2 is the control modulation wave obtained during the second frequency sweep, V p2 、V n2 The voltage obtained during the second frequency sweep; CV pp is the positive sequence control modulation wave transfer function corresponding to the positive sequence voltage, CV pn is the positive sequence control modulation wave transfer function corresponding to the negative sequence voltage, CV np is the negative sequence control modulation wave transfer function corresponding to the positive sequence voltage, CV nn is the negative sequence control modulation wave transfer function corresponding to the negative sequence voltage;

[0016]

[0017] In the formula, CI pp is the positive sequence control modulation wave transfer function corresponding to the positive sequence current, CI pn is the positive sequence control modulation wave transfer function corresponding to the negative sequence current, CI np is the negative sequence control modulation wave transfer function corresponding to the positive sequence current, CI nn is the negative sequence control modulation wave transfer function corresponding to the negative sequence current;

[0018] Control information CV pn Extract the control modulation wave c acabc With three-phase voltage u acabc The transfer function, control information CI pn Extract the control modulation wave c acabc With the three-phase current i acabcThe transfer function of

[0019] Step 2: Controller information neural network training

[0020] First stage training: The neural network is a 5-input 16-output system with inputs of frequency f and d-axis voltage V d , q-axis voltage V q , d-axis current I d , q-axis current I q , the output is the control information CV pp 、CV pn 、CV np 、CV nn , CI pp , CI pn , CI np , CI nn The amplitude and phase angle of

[0021] Second stage training: The neural network is a 21-input 8-output system. The controller information CV trained in the first stage pn and CI pn The training results and frequency f, d-axis voltage V d , q-axis voltage V q , d-axis current I d , q-axis current I q Together they serve as the input of the second-stage neural network, and the output is the sequence impedance Z pp , Z pn , Z np , Z nn Amplitude and phase angle.

[0022] Compared with the prior art, the present invention has the following advantages:

[0023] 1. The control information proposed in the present invention is compared with the frequency f, the working point (d-axis voltage V d ,q-axis voltage V q ,d-axis current I d ,q-axis current I q ), which is more closely related to the impedance characteristics and can optimize the neural network input data.

[0024] 2. The present invention considers using the control information existing in the neural network impedance modeling, and introduces the control information into the impedance modeling process, which can improve the accuracy of the neural network modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is the voltage source converter topology and control structure;

[0026] Figure 2 It is a schematic diagram of the control block diagram;

[0027] Figure 3 This is the structure diagram of the controller information neural network;

[0028] Figure 4 CV for controller information pn Sweep results;

[0029] Figure 5 CV for controller information pn Modeling results;

[0030] Figure 6 CV for controller information pn Modeling errors;

[0031] Figure 7 Controller information CI pn Sweep results;

[0032] Figure 8 Controller information CI pn Modeling results;

[0033] Fig. 9 Controller information CI pn Modeling errors;

[0034] Fig.10 is the impedance sweep result;

[0035] Fig.11 Modeling results for impedance;

[0036] Fig.12 Modeling error for impedance. DETAILED DESCRIPTION

[0037] The technical solution of the present invention is further described below in conjunction with the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be included in the protection scope of the present invention.

[0038] The present invention provides a neural network impedance modeling method using controller information, provides a method for using controller information, and provides a neural network training method for controller information. The method specifically includes the following steps:

[0039] Step 1: Control Information Modeling

[0040] The voltage source converter topology corresponding to the neural network based on controller information is as follows Figure 1 As shown. The voltage source converter adopts inner loop current decoupling control, and the system works in the inverter state. The inner loop current control gives the d-axis reference value i dref With q-axis reference value i qref, the grid phase angle θ is tracked by a phase locked loop (PLL). dc 、i dc Represent the DC bus voltage and DC current respectively, N is the grid voltage reference point, ω 1 is the fundamental frequency; u acabc with i acabc Indicates the three-phase AC output voltage and current of a, b, and c, u gabc Indicates the three-phase voltage of the power grid; R f and L f are AC equivalent resistance and AC equivalent inductance respectively; i d 、u d 、i q 、u q are the d-axis current, voltage and q-axis current, voltage respectively; K d is the decoupling coefficient between d and q currents, m dm To generate the control modulation wave.

[0041] The sequence impedance of the voltage source converter can be obtained by injecting a certain frequency of three-phase positive sequence voltage disturbance and negative sequence voltage disturbance signals on the AC side of the voltage source converter, and using Fourier analysis to analyze the current I of the first group respectively. p1 ,I n1 , voltage V p1 、V n1 , the current of the second group I p2 ,I n2 , voltage V p2 、V n2 . Sequence impedance Z pn The calculation formula is:

[0042]

[0043] In the formula, I p1 ,I n1 is the current obtained during the first group of frequency sweeps, V p1 、V n1 is the voltage obtained during the first group of frequency sweeps; I p2 ,I n2 is the current obtained during the second frequency sweep, V p2 、V n2 is the voltage obtained during the second group of frequency sweep; Z pp is the positive sequence voltage transfer function corresponding to the positive sequence current, Z pn is the positive sequence voltage transfer function corresponding to the negative sequence current, Z np is the negative sequence voltage transfer function corresponding to the positive sequence current, Z nn is the negative sequence voltage transfer function corresponding to the negative sequence current.

[0044] Controller Information CV pn and CI pn The acquisition method of is the same as that of impedance, except that two sets of control modulation waves are required, namely C p1 , C n1 and C p2 , C n2 , C p1 , C n1 , C p2 , C n2 Also obtained through Fourier analysis, the control information CV pn and CI pn Defined as:

[0045]

[0046] In the formula, C p1 , C n1 is the control modulation wave obtained during the first group of frequency sweep, V p1 、V n1 is the voltage obtained during the first group of frequency sweeps; C p2 , C n2 is the control modulation wave obtained during the second frequency sweep, V p2 、V n2 The voltage obtained during the second frequency sweep; CV pp is the positive sequence control modulation wave transfer function corresponding to the positive sequence voltage, CV pn is the positive sequence control modulation wave transfer function corresponding to the negative sequence voltage, CV np is the negative sequence control modulation wave transfer function corresponding to the positive sequence voltage, CV nn It is the negative sequence control modulation wave transfer function corresponding to the negative sequence voltage.

[0047]

[0048] In the formula, C p1 , C n1 is the control modulation wave obtained during the first group of frequency sweep, I p1 ,I n1 is the current obtained during the first group of frequency sweeps; C p2 , C n2 is the control modulation wave obtained during the second frequency sweep, I p2 ,I n2 The current obtained during the second group of frequency sweep; CI pp is the positive sequence control modulation wave transfer function corresponding to the positive sequence current, CI pn is the positive sequence control modulation wave transfer function corresponding to the negative sequence current, CI np is the negative sequence control modulation wave transfer function corresponding to the positive sequence current, CI nnIt is the negative sequence control modulation wave transfer function corresponding to the negative sequence current.

[0049] The control loop of a general voltage source converter can be simplified as follows: Figure 2 As shown in the control block diagram, the sequence impedance Z pn The displayed value is v acabc with i acabc The transfer function of the control information CV proposed by the present invention is pn and CI pn It reflects the control influence of the voltage source converter, the control information CV pn Extract the control modulation wave c acabc With three-phase voltage u acabc The transfer function, control information CI pn Extract the control modulation wave c acabc With the three-phase current i acabc The transfer function makes full use of the controller information that can be obtained.

[0050] In this step, the series voltage method is used to inject disturbances for frequency sweeping. Two sets of three-phase disturbance voltages are required, namely v inj1 、v inj2 :

[0051]

[0052] Where V inj is the amplitude of the injected voltage disturbance, ω inj is the injection frequency, ω 1 is the base frequency.

[0053] In this step, obtain the controller information CV pn , CI pn With sequence impedance Z pn The specific steps are:

[0054] 1. Inject the first set of voltage disturbances v inj1 The measured phase voltage is v p1 (ω inj ),v n1 (ω inj ), the phase current is i p1 (ω inj ),i n1 (ω inj ), control the modulation wave to be c p1 (ω inj ),c n1 (ω inj ) and save.

[0055] 2. Inject the second set of voltage disturbances v inj2 The measured phase voltage is v p2(ω inj ),v n2 (ω inj ), the phase current is i p2 (ω inj ),i n2 (ω inj ), control the modulation wave to be c p2 (ω inj ),c n2 (ω inj ) and save.

[0056] 3. According to the measured phase voltage and phase current, use the impedance calculation formula (1) to calculate the sequence impedance Z pn ; According to the measured phase voltage and control modulation wave, the controller information CV is calculated using formula (2) pn ; According to the measured phase current and control modulation wave, the controller information CI is calculated using formula (3) pn .

[0057] 4. Change the injection frequency ω inj , repeat the above frequency sweep steps to calculate different injection frequencies ω inj Controller information under CV pn , CI pn With sequence impedance Z pn .

[0058] Step 2: Controller information neural network training

[0059] According to the characteristics of impedance continuity, a multi-layer feedforward neural network of back propagation (BP) is selected for impedance identification. The controller information neural network proposed in the present invention makes full use of the controller information c that can be obtained. abc , which is used as the controller input of the neural network.

[0060] The schematic diagram of the controller information neural network training method proposed by the present invention is as follows: Figure 3 As shown, it is divided into two stages. The main idea is: pn and CI pn Introduce, use controller information CV pn and CI pn Help neural network training system sequence impedance Z pn , rather than being directly based on the frequency f, the operating point (d-axis voltage V d ,q-axis voltage V q ,d-axis current I d ,q-axis current I q ) Using neural network training sequence impedance Z pnThe controller information neural network training method proposed in the present invention selects the output (sequence impedance Z pn ) Closer input (controller information CV pn and CI pn ), which is beneficial to neural network training.

[0061] The neural network in the first stage is a 5-input 16-output system, with the inputs being the frequency f, the operating point (d-axis voltage V d ,q-axis voltage V q ,d-axis current I d ,q-axis current I q ), the output is the control information CV pp 、CV pn 、CV np 、CV nn , CI pp , CI pn , CI np , CI nn The amplitude (such as CV pp _ mag ) and phase angle (e.g. CV pp _ ang ). Different from the general neural network training method, the controller information neural network training method proposed in the present invention first obtains the controller information CV by training pn and CI pn , controller information CV pn and CI pn As the available information, the training results are verified. In the second stage, the trained controller information CV pn and CI pn as system input.

[0062] The neural network in the second stage is a 21-input 8-output system. The controller information CV trained in the first stage pn and CI pn The training results are related to the frequency f, the working point (d-axis voltage V d ,q-axis voltage V q ,d-axis current I d ,q-axis current I q ) are used as the input of the second-stage neural network, and the output is the sequence impedance Z pp , Z pn , Z np , Z nn The magnitude (e.g. Z pp _ mag ) and phase angle (e.g. Z pp _ ang ).

[0063] During the training process, 80% of the data is used for neural network training, 10% of the data is used for neural network testing, and 10% of the data is used for neural network verification. The neural network used contains 1 input layer, 3 hidden layers, and 1 output layer. The number of neurons in the hidden layer is determined by experiments.

[0064] In this step, the controller information CV is used pn , CI pn As an additional input to train the impedance network, in the first stage the controller information CV is trained pn , CI pn When 16 groups of controller information CV pn , CI pn One-time training can also be used to train 8 sets of controller information CV pn With 8 groups of controller information CI pn Train separately to reduce CV pn With CI pn At the same time, the interference caused by training; for the second layer of neural network input data, frequency f, working point (d-axis voltage V d ,q-axis voltage V q ,d-axis current I d ,q-axis current I q ) as a required input, for controller information, in addition to the controller information CV pn , CI pn In addition to using all as input, you can also use a single controller information CV pn or controller information CI pn As input, the selection principle is the controller information CV pn , CI pn With sequence impedance Z pn degree of similarity.

[0065] 1. Controller information modeling effect

[0066] In order to verify the above-mentioned neural network voltage source converter impedance modeling method based on controller information, three-phase positive sequence voltage disturbance and negative sequence voltage disturbance signals with a frequency range of 1 to 200 Hz and an interval of 2 Hz are injected into the AC side of the voltage source converter. d =10A to i d =20A, the interval is 2A.

[0067] Figure 4 , Figure 5 , Figure 6 Respectively, controller information CV pn Sweep results, controller information CV pn Modeling results, controller information CV pn Modeling error. Figure 7 , Figure 8 , Fig. 9 Controller information CI pn Sweep results, controller information CI pn Modeling results, controller information CI pn Modeling error. At this time, the controller information CV pn The training error is within 5%, and the controller information CI pn The training error is within 3%, but it does not affect the accuracy of the impedance modeling results of the method proposed in the present invention. The controller information training result CV pn and CI pn As the input of the second stage neural network.

[0068] 2. Impedance modeling effect

[0069] The controller information training result CV through the first stage pn and CI pn Training sequence impedance. Fig.10 , Fig.11 , Fig.12 They are impedance sweep results, impedance modeling results, and impedance modeling errors. d ,q-axis voltage V q ,d-axis current I d ,q-axis current I q ) and sequence impedance Z pn Not an obvious relationship, controller information CV pn and CI pn With sequence impedance Z pn The characteristics are very similar. The impedance modeling method of the controller information proposed in the present invention makes full use of the controller information CV pn and CI pn With sequence impedance Z pn The relationship between the impedance training results has a smaller error.

[0070] At this time, the impedance modeling error is within 0.5%, the average impedance amplitude error is 0.23dB, the maximum error is 0.26dB, the average impedance phase angle error is 0.31°, and the maximum error is 0.41°. The accuracy of the method proposed in the present invention is verified.

[0071] according to Figure 4 , Figure 7 , Fig.10 Sweep results, controller information CI pn With sequence impedance Z pn More similar, you can also select only the controller information CI pn As the controller input of the second-stage neural network, it improves the computing speed of the neural network.

Claims

1. A neural network impedance modeling method using controller information, characterized in that The method comprises the following steps: Step 1: Control Information Modeling Step 11: Obtain the sequence impedance Z of the voltage source converter pn Sequence impedance Z pn The calculation formula is: In the formula, I p1 ,I n1 is the current obtained during the first group of frequency sweeps, V p1 、V n1 is the voltage obtained during the first group of frequency sweeps; I p2 ,I n2 is the current obtained during the second frequency sweep, V p2 、V n2 is the voltage obtained during the second group of frequency sweep; Z pp is the positive sequence voltage transfer function corresponding to the positive sequence current, Z pn is the positive sequence voltage transfer function corresponding to the negative sequence current, Z np is the negative sequence voltage transfer function corresponding to the positive sequence current, Z nn is the negative sequence voltage transfer function corresponding to the negative sequence current; Step 12: Get controller information CV pn and CI pn Control information CV pn and CI pn Defined as: In the formula, C p1 , C n1 is the control modulation wave obtained during the first group of frequency sweep, V p1 、V n1 is the voltage obtained during the first group of frequency sweeps; C p2 , C n2 is the control modulation wave obtained during the second frequency sweep, V p2 、V n2 The voltage obtained during the second frequency sweep; CV pp is the positive sequence control modulation wave transfer function corresponding to the positive sequence voltage, CV pn is the positive sequence control modulation wave transfer function corresponding to the negative sequence voltage, CV np is the negative sequence control modulation wave transfer function corresponding to the positive sequence voltage, CV nn is the negative sequence control modulation wave transfer function corresponding to the negative sequence voltage; In the formula, CI pp is the positive sequence control modulation wave transfer function corresponding to the positive sequence current, CI pn is the positive sequence control modulation wave transfer function corresponding to the negative sequence current, CI np is the negative sequence control modulation wave transfer function corresponding to the positive sequence current, CI nn is the negative sequence control modulation wave transfer function corresponding to the negative sequence current; Control information CV pn Extract the control modulation wave c acabc With three-phase voltage u acabc The transfer function, control information CI pn Extract the control modulation wave c acabc With the three-phase current i acabc The transfer function of Step 2: Controller information neural network training First stage training: The neural network is a 5-input 16-output system with inputs of frequency f and d-axis voltage V d , q-axis voltage V q , d-axis current I d , q-axis current I q , the output is the control information CV pp 、CV pn 、CV np 、CV nn , CI pp , CI pn , CI np , CI nn The magnitude and phase angle of Second stage training: The neural network is a 21-input 8-output system. The controller information CV trained in the first stage pn and CI pn The training results and frequency f, d-axis voltage V d , q-axis voltage V q , d-axis current I d , q-axis current I q Together they serve as the input of the second-stage neural network, and the output is the sequence impedance Z pp , Z pn , Z np , Z nn Amplitude and phase angle.

2. The neural network impedance modeling method using controller information according to claim 1 is characterized in that In step 1, obtain controller information CV pn , CI pn With sequence impedance Z pn The specific steps are: (1) Inject the first set of voltage disturbances v inj1 The measured phase voltage is v p1 (ω inj ),v n1 (ω inj ), the phase current is i p1 (ω inj ),i n1 (ω inj ), control the modulation wave to be c p1 (ω inj ),c n1 (ω inj ) and save; (2) Inject the second set of voltage disturbances v inj2 The measured phase voltage is v p2 (ω inj ),v n2 (ω inj ), the phase current is i p2 (ω inj ),i n2 (ω inj ), control the modulation wave to be c p2 (ω inj ),c n2 (ω inj ) and save; (3) Based on the measured phase voltage and phase current, the sequence impedance Z is calculated pn ; According to the measured phase voltage and control modulation wave, the controller information CV is calculated pn ; According to the measured phase current and control modulation wave, the controller information CI is calculated pn ; (4) Change the injection frequency ω inj , repeat the above frequency sweep steps to calculate different injection frequencies ω inj Controller information under CV pn , CI pn With sequence impedance Z pn .

Images