Converter adaptive hybrid synchronization control method based on steady-state margin prediction

Abstract

The application discloses a converter adaptive hybrid synchronization control method based on steady-state margin prediction, belongs to the technical field of new energy grid connection, and comprises the following steps: constructing a grid-connected VSC converter adaptive hybrid synchronization control dq impedance model; constructing a steady-state margin prediction model based on DNN; collecting training data; training the steady-state margin prediction model in step S2 by using the collected training data to obtain a trained steady-state margin prediction model; based on the steady-state margin prediction model in step S3, solving the optimal solution of adaptive coefficients by using an interior point method under the condition of a given steady-state margin; and based on the optimal solution of adaptive coefficients obtained by combining the steady-state margin prediction model in step S3 and the optimization algorithm in step S4, adjusting parameters to realize adaptive small-disturbance stability control of the converter. k a The application discloses a converter adaptive hybrid synchronization control method based on steady-state margin prediction, belongs to the technical field of new energy grid connection, and comprises the following steps: constructing a grid-connected VSC converter adaptive hybrid synchronization control dq impedance model; constructing a steady-state margin prediction model based on DNN; collecting training data; training the steady-state margin prediction model in step S2 by using the collected training data to obtain a trained steady-state margin prediction model; based on the steady-state margin prediction model in step S3, solving the optimal solution of adaptive coefficients by using an interior point method under the condition of a given steady-state margin; and based on the optimal solution of adaptive coefficients obtained by combining the steady-state margin prediction model in step S3 and the optimization algorithm in step S4, adjusting parameters to realize adaptive small-disturbance stability control of the converter. k a The application discloses a converter adaptive hybrid synchronization control method based on steady-state margin prediction, belongs to the technical field of new energy grid connection, and comprises the following steps: constructing a grid-connected VSC converter adaptive hybrid synchronization control dq impedance model; constructing a steady-state margin prediction model based on DNN; collecting training data; training the steady-state margin prediction model in step S2 by using the collected training data to obtain a trained steady-state margin prediction model; based

Description

TECHNICAL FIELD

[0001] The present application relates to the field of new energy grid-connection technology, and particularly relates to a converter adaptive hybrid synchronization control method based on steady-state margin prediction. BACKGROUND

[0002] As the interface between new energy and power grid, the stability of the converter will directly affect the reliability of the power grid operation. In the current new energy grid-connection control technology, the converter control strategy can be roughly divided into two types: grid-following control and grid-forming control. The main difference between the two is the way of achieving grid voltage synchronization. Among them, the grid-following converter tracks the grid information through the phase-locked loop; the grid-forming control achieves synchronization function through the power synchronization loop which simulates the characteristics of synchronous generators. The difference in synchronization methods will lead to significant differences in the oscillation instability characteristics of the two control strategies. Specifically, the grid-following inverter will lose stability under weak grid conditions due to the failure of the phase-locked loop to track the grid information; the grid-forming inverter is not easy to synchronize under strong grid due to the lack of damping; in addition, the stability performance of the two in terms of renewable energy output also exhibits a dual characteristic. Therefore, under the condition of rapid and large fluctuations of renewable energy and power load, it is difficult to maintain the stability of the grid-converter system by using a single control strategy, and the power system urgently needs a new type of converter control strategy.

[0003] The adaptive hybrid synchronization control is a control strategy that integrates the phase-locked loop into the power synchronization loop, combining the characteristics of grid-following inverters and grid-forming inverters, and can adjust the adaptive coefficient k a to achieve output proportion adjustment. However, the existing adaptive hybrid synchronization control only adjusts the parameters according to the short-circuit ratio, ignoring the stability influence of the fluctuation of new energy itself; on the other hand, the existing short-circuit ratio parameter adjustment method only gives the linear variation of the adaptive coefficient with the short-circuit ratio, while the actual steady-state margin and the short-circuit ratio have nonlinear characteristics, and are related to the composition of each control loop of the grid-forming VSC converter, so the original parameter adjustment method cannot well adapt to the actual steady-state margin variation relationship, and also lacks universality. In addition, this parameter adjustment method can only qualitatively improve the small signal stability of the system, and cannot accurately and quantitatively control the steady-state margin. In fact, although adjusting the adaptive coefficient k a will improve the small signal stability of the grid-converter system, it will also reduce the proportion coefficient of the power loop and thus reduce the active power dynamic tracking performance of the converter, and a higher adaptive coefficient will also cause the grid-forming converter to lose stability monotonously under weak grid conditions.

[0004] Therefore, how to provide an adjustable adaptive coefficient parameter adjustment method with universality and high accuracy is an urgent technical problem for those skilled in the art. SUMMARY

[0005] To this end, the application provides a converter adaptive hybrid synchronization control method based on steady-state margin prediction to solve the problems in the prior art.

[0006] To achieve the above-mentioned purpose, the application provides the following technical solutions.

[0007] The converter adaptive hybrid synchronization control method based on steady-state margin prediction comprises the following steps:

[0008] Step S1: constructing a grid-forming VSC converter adaptive hybrid synchronization control dq impedance model;

[0009] Step S2: constructing a steady-state margin prediction model based on deep neural networks (DNN);

[0010] Step S3: collecting training data, combining the dq impedance model established in step S1 to calculate the steady-state margin of the converter under the working condition corresponding to the training data as the output label of the steady-state margin prediction model in step S2, training the steady-state margin prediction model in step S2 using the collected training data, and obtaining the trained steady-state margin prediction model;

[0011] Step S4: based on the steady-state margin prediction model in step S3, using the interior point method to solve the optimal solution of the adaptive coefficient k a under the condition of a given steady-state margin;

[0012] Step S5: based on the steady-state margin prediction model in step S3, inputting the short-circuit ratio and the reference value of active power of the power grid in real time to generate a dynamic prediction result of the steady-state margin of the converter, and combining the optimization method of step S4 to obtain the optimal solution of the adaptive coefficient k a for parameter adjustment, realizing adaptive small disturbance stability control of the converter.

[0013] Further, in step S1, the method for constructing the grid-forming VSC converter adaptive hybrid synchronization control dq impedance model comprises the following steps:

[0014] Step S101: building a physical model of the converter adaptive hybrid synchronization control;

[0015] Step S102: On the basis of considering the dynamic process of the phase-locked loop and the power synchronization link, a small signal dq impedance modeling method is used to construct a small signal equation of the hybrid synchronization link, the reactive power-voltage droop control link, the voltage control link, the current control link, the coordinate transformation link and the physical circuit according to the physical model;

[0016] Step S103: The small signal equation constructed in step S102 is solved to obtain a converter dq impedance transfer function.

[0017] Further, in the step S2, a method for constructing a DNN-based steady-state margin prediction model is specifically as follows:

[0018] Step S201: Analyze the key factors affecting the small signal stability of the power grid-converter system, screen the key features affecting the small signal stability of the power grid-converter system, and use feature engineering to construct the input label X of the steady-state margin prediction model I ;

[0019] Step S202: Determine the DNN structure of the steady-state margin prediction model by considering the number of input labels and output labels, and select a loss function L for optimizing the weights and biases of the steady-state margin prediction model in step S3;

[0020] Step S203: The steady-state margin solved in step S1 is taken as the output label Y I , and the input label X I in step S201 is combined to construct a sample data set D I ={X I , Y I}, which is used for regression training of the steady-state margin prediction model in step S3.

[0021] Further, the input label X I of the steady-state margin prediction model in step S201 is represented as:

[0022] ;

[0023] Wherein, k aI is the converter adaptive coefficient in the sample data, P refI is the converter operating condition characteristic active power reference value in the sample data, SCR I is the grid-side short circuit ratio in the sample data.

[0024] Further, the DNN structure of the steady-state margin prediction model in step S202 is a four-layer structure, specifically including an input layer, an output layer and two hidden layers, wherein the activation function type of the hidden layer neurons is selected as SigmoidThe activation function of the input layer neurons is a sigmoid function, while the activation function of the output layer neurons is a linear function. The output of each layer neuron is represented by the following formula:

[0025] ;

[0026] ;

[0027] where w and b represent the weight and bias, respectively; n represents the layer number of the neuron, and the input layer is defined as the 0th layer; p represents the output of the n th neuron in the p -1 layer; q represents the output of the n th neuron in the q th layer; k represents the output of the n th neuron in the m -1 layer; n represents the output of the h th neuron in the n th layer; q represents the output of the q th neuron in the w th hidden layer; p,q represents the weight between the n th neuron in the p -1 layer and the n th neuron in the q th layer; x p is the output of the n th neuron in the p -1 layer; b q is the bias between the bias node in the n -1 layer and the q th neuron in the f th layer; I represents the output of the output layer neuron.

[0028] Further, the loss function L in the step S202 is represented as:

[0029] ;

[0030] where MSE is the mean square error, N represents the number of samples, I represents the Ith sample data, Y I is the steady-state margin solved in the step S1, f I represents the output of the output layer neuron.

[0031] Further, the output label YI is represented as:

[0032] ;

[0033] wherein GM I is the steady-state margin in the sample data.

[0034] Further, in the step S4, the adaptive coefficient k a is solved, and the method for solving the optimal solution specifically comprises the following steps:

[0035] Step S401: determining the steady-state margin reference value GM ref of the power grid-converter system a , and designing the optimization objective function F(k ref ) as the square difference between the steady-state margin predicted value in the step S2 and the steady-state margin reference value GM I , and the calculation formula is:

[0036] ;

[0037] wherein, f (X I ) represents the output of the output layer neuron, specifically the predicted value of the steady-state margin prediction model in the step S2, and X I is the corresponding input label;

[0038] Step S402: setting the optimization initial value k a of the converter adaptive coefficient k a0 , and setting the upper and lower threshold values, which correspond to the inequality constraints of the optimal problem;

[0039] Step S403: setting the iteration number and the objective function threshold value, and solving the optimal solution of the converter adaptive coefficient k a under the working condition by using the interior point method.

[0040] Further, the training data in the step S3 comprises the short-circuit ratio of the power grid side, the reference value of the active power, the PCC point voltage dq-axis component, the PCC point current dq-axis component, the converter bridge arm voltage dq-axis component and the converter bridge arm current dq-axis component.

[0041] The present application has the following advantages:

[0042] 1. The network type VSC converter adaptive hybrid synchronization control dq impedance model constructed in the present application considers the dynamic influence of the phase-locked loop and the power synchronization link on the control angle θ, deduces the transformation equation of the electrical quantity under the system coordinate system and the control coordinate system, and obtains the accurate full-order impedance transfer function expression, thereby improving the calculation accuracy of the steady-state margin.

[0043] 2. The DNN-based steady-state margin prediction model constructed by the application considers the influence of new energy station output variation on the steady-state margin on the basis of considering short-circuit ratio variation, and the trained steady-state margin prediction model can reflect the nonlinear characteristics between system steady-state margin and converter multi-conditions, and can realize fast and accurate steady-state margin estimation of the converter in a wide range of conditions.

[0044] 3. The adaptive hybrid synchronization control method constructed by the application can realize quantitative control of the steady-state margin in the form of solving the adaptive coefficient k a from the existing steady-state margin prediction model, and can improve small disturbance stability or active power dynamic tracking performance according to the needs of the system. BRIEF DESCRIPTION OF DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the application or the technical solutions in the prior art, the drawings needed in the following embodiment or prior art description will be briefly introduced. Obviously, the drawings in the following description are only exemplary, and for those skilled in the art, other embodiments can be obtained from the provided drawings without creative labor.

[0046] The structures, proportions, sizes, etc. shown in the specification are only used to cooperate with the content disclosed in the specification, to be understood and read by those skilled in the art, and are not used to limit the implementation conditions of the application, so they do not have technical significance. Any modification of the structure, change of the proportion relationship or adjustment of the size, without affecting the effect and purpose that can be achieved by the application, should still fall within the scope of the technical content disclosed by the application.

[0047] Figure 1 Flowchart of the adaptive hybrid synchronization control method for the grid-connected VSC converter based on the DNN steady-state margin prediction of the application;

[0048] Figure 2 LC grid-connected converter topology and adaptive hybrid synchronization control block diagram implemented by the application;

[0049] Figure 3 Full-order small-signal model diagram of the adaptive hybrid synchronization control for the grid-connected VSC converter constructed by the application;

[0050] Figure 4 Structure diagram of the DNN-based steady-state margin prediction model constructed by the application;

[0051] Figure 5 Schematic diagram of the training process of the DNN-based steady-state margin prediction model constructed by the application;

[0052] Figure 6Schematic diagram of the optimized parameter adjustment based on the DNN steady-state margin prediction model constructed in the present invention. DETAILED DESCRIPTION

[0053] The following describes the implementation of the present invention using specific embodiments. Those skilled in the art will readily understand the other advantages and benefits of the present invention from the disclosure herein. Obviously, the embodiments described are only a portion of the present invention, not all of it. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are intended to fall within the scope of protection of the present invention.

[0054] like Figure 1 As shown in FIG, the adaptive hybrid synchronous control method of the grid-type VSC converter based on DNN steady-state margin prediction includes the following steps:

[0055] Step S1: Constructing a dq impedance model for adaptive hybrid synchronous control of a grid-connected VSC converter, the purpose of which is to calculate the steady-state margin of the grid-converter system using the generalized Nyquist theorem in combination with grid-side impedance information under different converter operating conditions;

[0056] Step S2: Constructing a steady-state margin prediction model based on deep neural networks (DNN) to achieve steady-state margin prediction of the converter under multiple operating conditions;

[0057] Step S3: Collect training data, and calculate the steady-state margin of the converter under the working conditions corresponding to the training data in combination with the dq impedance model established in step S1. The calculated steady-state margin is used as the output label of the steady-state margin prediction model in step S2. The steady-state margin prediction model in step S2 is trained using the collected training data to obtain a trained steady-state margin prediction model.

[0058] Step S4: Based on the steady-state margin prediction model in step S3, the adaptive coefficient is solved using the interior point method under the condition of a given steady-state margin. k a The optimal solution of

[0059] Step S5: Based on the steady-state margin prediction model in step S3, the short-circuit ratio and active power reference values ​​of the real-time input grid are used to generate the dynamic prediction result of the converter steady-state margin, and the adaptive coefficient is obtained by combining the optimization algorithm of step S4. k a The optimal solution is used to adjust the parameters and realize the adaptive small disturbance stable control of the converter.

[0060] like Figure 2 The LC grid-connected converter topology and adaptive hybrid synchronous control block diagram are shown. In the topology, the converter is connected to the grid through a filter inductor. Lf and filter capacitor C f connected to the power grid, which is represented by an infinite power voltage source V g and equivalent inductance L g in series. The adaptive hybrid synchronization control structure is composed of a hybrid synchronization loop, a reactive-voltage droop control loop, a voltage control loop and a current control loop. The hybrid synchronization loop is composed of a phase-locked loop and a power synchronization loop. Figure 2 It can be seen that the hybrid synchronization loop is composed of a phase-locked loop and a power synchronization loop, and the phase angle signal of the converter is obtained by summing the power feedback signal and the q-axis voltage feedback signal of the PCC point:

[0061] ;

[0062] In the formula, k a is an adaptive coefficient; k p and k PLL are the active-frequency (P-F) proportional coefficient of the power synchronization loop controller and the proportional coefficient of the phase-locked loop controller, respectively; P f ; k ip is the integral control coefficient of the power synchronization loop controller; v oq is the q-axis component of the PCC voltage measurement value; q 0 is the rated angular frequency of the power grid; s represents the variable in the system coordinate system; ω is the reference value of the active power of the converter; P* LPF is the measured value of the active power processed by a first-order low-pass filter, and the calculation formula is: P

[0063] ;

[0064] In the formula, P is the measured value of the active power of the converter; ω LPF is the cutoff frequency of the first-order low-pass filter.

[0065] The reactive-voltage droop control loop generates the PCC point d axis voltage reference value v od * :

[0066] ;

[0067] In the formula, k q ​​is the droop coefficient of the reactive power-voltage loop controller; Q* is the reference value of the converter reactive power after being processed by a first-order low-pass filter LPF Q LPF ; v 0 is the preset amplitude of the PCC point voltage.

[0068] In the voltage control loop, the q-axis voltage reference value v oq * is generally set to 0. The voltage reference value generates the bridge arm voltage reference value v cd * and v cq * :

[0069] ;

[0070] ;

[0071] wherein: i cd *, i cq *, i cd , i cq are the reference value and the measured value of the bridge arm current d axis component and q axis component, respectively; v od , v oq are the q-axis component and the d-axis component of the PCC voltage measured value, respectively; d q k pv and k iv are the proportional coefficient and the integral coefficient of the voltage loop; k pi and k ii are the proportional coefficient and the integral coefficient of the current loop.

[0072] Compared with the general hybrid synchronization control structure, the control strategy introduces an adaptive coefficient k a to adjust the output ratio of the phase-locked loop and the power synchronization loop in the hybrid synchronization link, so as to achieve the purpose of affecting the steady-state margin of the system, and the total value of the output of the hybrid synchronization link is 1.

[0073] ​​In step S1, the method for constructing a dq impedance model for adaptive hybrid synchronous control of a grid-type VSC converter specifically includes the following steps:

[0074] Step S101: building a physical model of a converter with adaptive hybrid synchronous control;

[0075] Step S102: Based on the dynamic process of the phase-locked loop and power synchronization link, a small-signal dq impedance modeling method is used according to the physical model to construct small-signal equations for the hybrid synchronization link, the reactive-voltage droop control link, the voltage control link, the current control link, the coordinate transformation link, and the physical circuit;

[0076] Step S103: The small signal equations established in step S102 are simultaneously eliminated to obtain the converter dq impedance transfer function.

[0077] like Figure 3 The full-order small signal model of the grid-type VSC converter adaptive hybrid synchronous control constructed in step S1 shown in the figure, according to the impedance definition, the impedance transfer function can be expressed as:

[0078] ;

[0079] Among them, Z vsc ( s ) represents the full-order small signal model of the adaptive hybrid synchronous control of the grid-type VSC converter. The specific expressions of the transfer function matrices corresponding to this model are as follows:

[0080] ;

[0081] Among them, Z g , Z f 、Y c Derived from the physical circuit small signal equation, they represent the grid equivalent inductance impedance matrix, filter inductance impedance matrix, and filter capacitor admittance matrix respectively; G vo-vo , G vo-io , G io-vo , G io-io , G ic-vo , G ic-io , G vc-vo , G vc-io Derived from the small signal equations of the hybrid synchronous control link and the coordinate transformation link, they represent the PCC point voltage transformation voltage component matrix, PCC point voltage transformation current component matrix, grid side current transformation voltage component matrix, grid side current transformation current component matrix, bridge arm current transformation voltage component matrix, bridge arm current transformation current component matrix, bridge arm voltage transformation voltage component matrix, and bridge arm voltage transformation current component matrix respectively; G cv , G devThe voltage control loop small-signal equation is derived, and represents the voltage loop PI control matrix and the filter capacitor coupling matrix, respectively; G ci , dei The current control loop small-signal equation is derived, and represents the current loop PI control matrix and the filter inductor coupling matrix, respectively; G QI , QV The reactive-voltage droop control loop small-signal equation is derived, and represents the reactive droop current component matrix and the reactive droop voltage component matrix, respectively; I od , I oq are the d axis component and the q axis component of the bridge arm current steady-state value, respectively; I cd , I cq are the d axis component and the q axis component of the PCC point current steady-state value, respectively; V cd , V cq are the d axis component and the q axis component of the bridge arm voltage steady-state value, respectively; V od , V oq are the d axis component and the q axis component of the PCC point voltage steady-state value, respectively; in addition, s represents the variable in the system coordinate system, and the symbol "Δ" represents the small-signal component, i o is the preset amplitude of the grid-side current, i od , i oq are the d axis component and the q axis component of the grid-side current, respectively.

[0082] Then, the grid-side impedance model can be expressed as:

[0083] ;

[0084] In addition, according to the generalized Nyquist stability criterion, the stability of the grid-converter system can be predicted by the characteristic locus. The characteristic root locus is the characteristic value locus of the impedance ratio parameterized as a function of frequency, which can be expressed as:

[0085] ;

[0086] In the formula,I_2 represents the 2×2 identity matrix, λ It's Z vsc ( s ) and Z g ( s ), det is the matrix determinant.

[0087] Moreover, the steady-state margin of the system can be expressed by the amplitude margin on the Nyquist diagram, and its specific expression is:

[0088] ;

[0089] Where GM is the steady-state margin of the grid-converter system, ω c is the crossing frequency corresponding to the characteristic root locus, and j is an imaginary unit.

[0090] In step S2, a method for constructing a DNN steady-state margin prediction model is constructed, which specifically includes the following steps:

[0091] Step S201: Analyze the key factors affecting the small-disturbance stability of the grid-converter system, select the key features affecting the small-disturbance stability of the grid-converter system, and use feature engineering to construct the input label X of the steady-state margin prediction model. I ;

[0092] Step S202: considering the number of input labels and output labels, determining the DNN structure of the steady-state margin prediction model, and selecting an appropriate loss function L to optimize the weights and biases of the steady-state margin prediction model in step S3;

[0093] Step S203: The steady-state margin obtained in step S1 is used as the output label Y I , combined with the input label X in S201 I Construct sample data set D I ={X I , Y I}, used for regression training of the steady-state margin prediction model in step S3.

[0094] In order to predict the steady-state margin, it is necessary to screen the input labels that affect the system steady-state margin. These features mainly include renewable energy output, grid-side short-circuit ratio ( SCR ) and the adaptive coefficient ( k a ) three indicators. Considering that new energy units usually only provide active power to the grid, the output of new energy can be calculated by the reference value of active power P ref Instead, input the sample set X of labels I It can be expressed as:

[0095] ;

[0096] in, k aI is the converter adaptive coefficient in the sample data, P refI is the reference value of the converter operating condition characteristic active power in the sample data, SCR I is the grid side short-circuit ratio in the sample data.

[0097] The structure of the steady-state margin prediction model based on DNN constructed in step S2 is as follows: Figure 4 The DNN model is a high-order nonlinear model with three inputs and one output, and has a four-layer structure consisting of an input layer X1-X3, an output layer Y1, and two hidden layers M1-M 11 and N1-N 11 The number of neurons in each layer is 3-11-11-1. The activation function type of the hidden layer neurons is selected as Sigmoid function, while the activation function of the output layer neurons is a linear function. The output of each layer of neurons can be expressed by the following formula:

[0098] ;

[0099] ;

[0100] Among them, w and b represent weight and bias respectively; n Represents the number of layers in which the neurons are located, and stipulates that the input layer is the 0th layer; p represent n -1st floor p neurons; q represent n Layer q neurons; k represent n -1 layer corresponds to the number of neurons; m represent n The number of neurons corresponding to the layer; h q Represents n The hidden layer q The output of a neuron; w p,q Represents n -1st floor p neurons and n Layer q The weights between neurons; x p yes n -1st floor p The output of a neuron; b q yesn -1 layer bias node and the bias between the n-th neuron in the n layer; q f (X I ) represents the output of the output layer neuron.

[0101] The steady-state margin prediction model is trained using the back propagation algorithm and the Levenberg-Marquardt method, and the mean square error (MSE) is used as the loss function for model training L , and the specific expression is:

[0102] ;

[0103] In the formula, N represents the number of samples, I represents the I-th sample data, Y I represents the steady-state margin calculated by the dq impedance model constructed by S1, and can be expressed as:

[0104] ;

[0105] Where, GM I is the steady-state margin in the sample data.

[0106] In step S3, the training data includes the short-circuit ratio of the power grid side, the reference value of the active power, the PCC point voltage dq axis component, the PCC point current dq axis component, the converter bridge arm voltage dq axis component, and the converter bridge arm current dq axis component. The process of collecting training data can be described as: using MATLAB / SINMULINK software to build the simulation model shown in Figure 2 , and combining MATLAB scripts to call MATLAB / SINMULINK for batch simulation. The active power reference value P ref is set to 0pu-1pu, the short-circuit ratio SCR is set to 2-20, and the adaptive coefficient k a is uniformly changed from 0 to 0.9, a total of 3200 groups of data are taken, and the training set, test set and validation set are divided according to the ratio of 3:1:1. The output label Y I of the steady-state margin prediction model will be calculated by importing the steady-state values and grid impedance of each operating point to the dq impedance model constructed in step S1 during training. The steady-state margin prediction model after training can be used for online prediction of steady-state margin.

[0107] The steady-state margin model training process is shown in Figure 5 , and the grid-connected converter measures the signal data set [P ref , SCR , k a ​, the dq impedance model of step S1 and the grid impedance to obtain a stability margin dataset [P ref , SCR , k a , the training dataset [P ref , SCR , k a train and the test dataset [P ref , SCR , k a test , the training dataset and the test dataset to train and test the neural network, adjust the number of model layers and nodes, obtain the mean square error MSE that reflects the model accuracy and model prediction ability, if the MSE is less than the threshold value MSE limit , continue to further process the neural network, otherwise, retrain the neural network.

[0108] In step S4, the adaptive coefficient k a is solved. Figure 6 As shown in the above training results, the neural network is further processed, the neural network is de-normalized, the steady state margin is evaluated, the steady state margin reference value GM ref and the optimal initial value k a0 are set, the interior point method is used for optimization, when the iteration process reaches the threshold value of the objective function or the corresponding iteration number, the iteration is stopped, the optimal solution of the adaptive coefficient k a is obtained, otherwise, the adaptive coefficient k a is iteratively processed, the input feature quantity measurement signal and the value of k a are taken, and the neural network is normalized and preprocessed.

[0109] In step S4, the adaptive coefficient k a is solved.

[0110] Step S401: determine the steady state margin reference value GM ref of the grid-converter system, design the optimization objective function F( k a ) as the square difference between the steady state margin prediction value of step S2 and the steady state margin reference value GM ref , and the calculation formula is:

[0111] ​​ ;

[0112] where f(X I ) represents the predicted value of the steady-state margin prediction model in step S2, X I is the corresponding input label;

[0113] Step S402: Set the adaptive coefficient of the converter k a Optimization initial value k a0 , and set the upper and lower threshold values corresponding to the inequality constraints of the optimization problem;

[0114] Step S403: Set the number of iterations and the threshold value of the objective function, and solve the adaptive coefficient of the converter k a The optimal solution under this working condition.

[0115] Since the DNN model is essentially a high-order nonlinear function, it is quite difficult to deduce the input label from the output label, and numerical optimization iteration method needs to be used for solving. First, define the objective function F ( k a ), which is the square difference between the predicted value f(X I ) of the steady-state margin prediction model and the steady-state margin reference value GM ref . In addition, in order to maintain the converter without losing control of active power, the adaptive coefficient k a needs to be constrained. Both of them constitute the constraints of the optimization problem, which is expressed as:

[0116] ;

[0117] For the above constraint conditions, the objective function and inequality constraint are converted into Lagrange function, and the interior point method is used for optimization, which is expressed as:

[0118] ;

[0119] In the formula, μ is the barrier parameter, which is used to control the strength of the constraint, and gradually decreases with the progress of the iteration process to gradually approach the corresponding constraint boundary. The update rule of the parameter is:

[0120] ;

[0121] In the formula, β is the reduction factor, which is used to gradually reduce the barrier parameter μ , r is a discrete index.

[0122] Optimization direction Δ of the interior point method α It can be determined by the gradient of the Lagrangian function and the Hessian matrix, and its specific expression is:

[0123] ;

[0124] Then, the adaptive coefficients are updated by Newton steps k a , the specific expression is:

[0125] ;

[0126] Where, α It is the iteration step size, which is automatically solved by the backtracking line search algorithm during the iteration process to ensure that the objective function value continues to decrease until the constraints are met.

[0127] Moreover, when the iterative process reaches the threshold of the objective function or the corresponding number of iterations, the iteration will stop. k a As variables, and the range of inequality constraints is small, the control system can quickly calculate the values ​​of the gradient and Hessian matrix, which can meet the needs of online calculation of the system.

[0128] The steady-state margin prediction model trained in step S5 can be used to calculate the adaptive coefficient based on the input characteristic data of the system. k a The optimal solution is adjusted online and the prediction results of the system's steady-state margin are generated in real time.

[0129] Although the present invention has been described in detail above using general descriptions and specific embodiments, it will be apparent to those skilled in the art that modifications and improvements may be made thereto. Therefore, such modifications and improvements, without departing from the spirit of the present invention, are intended to be within the scope of protection claimed herein.

Claims

1. A converter adaptive hybrid synchronous control method based on steady-state margin prediction, characterized in that, The method comprises the following steps: Step S1: constructing an adaptive hybrid synchronous control dq impedance model of a grid-connected VSC converter; Step S2: constructing a steady-state margin prediction model based on a DNN; Step S3: collecting training data, calculating the steady-state margin of the converter under the working condition corresponding to the training data in combination with the dq impedance model established in step S1, taking the steady-state margin prediction model in step S2 as the output label, training the steady-state margin prediction model in step S2 using the collected training data, and obtaining the trained steady-state margin prediction model; Step S4: Based on the steady-state margin prediction model in step S3, solve the adaptive coefficient k using the interior point method under the condition of given steady-state margin a optimal solution; Step S5: Based on the steady-state margin prediction model in step S3, the short-circuit ratio and the reference value of the active power of the power grid are input in real time to generate a dynamic prediction result of the converter steady-state margin, and the adaptive coefficient k is obtained by combining the optimization method in step S4 a The optimal solution is adjusted to realize adaptive small disturbance stability control of the converter; The adaptive hybrid synchronous control structure is composed of a hybrid synchronous ring, a reactive power-voltage droop control ring, a voltage control ring, and a current control ring. The hybrid synchronous ring is composed of a phase-locked loop and a power synchronous link. The phase angle signal of the converter is obtained by summing the power feedback signal and the q-axis voltage feedback signal of the PCC point: ； where k a is an adaptive coefficient; k p and k PLL are the proportional coefficients of the active- frequency (P-f) loop controller and the phase-locked loop (PLL) loop controller, respectively; k ip is the integral control coefficient of the P-f loop controller; v oq is the q-axis component of the PCC voltage measurement; ω0is the rated angular frequency of the grid; s denotes the variable in the system coordinate; P*is the reference value of the active power of the converter; P LPF is the active power measurement processed by a first-order low-pass filter. In the step S4, the adaptive coefficient k is solved a The method for obtaining the optimal solution specifically comprises the following steps. Step S401: Determine the steady-state margin reference value GM of the power grid-converter system ref The design optimization objective function F(k a ) is the square difference between the steady-state margin prediction value in step S2 and the steady-state margin reference value GM ref , and the calculation formula is: ； wherein f(X I ) represents the output of the output layer neuron, specifically the predicted value of the steady-state margin prediction model in step S2, X I is the corresponding input label; Step S402: setting the converter adaptive coefficient k a the optimal initial value k a0 and setting the upper and lower threshold values corresponding to the inequality constraints of the optimization problem; Step S403: Set the iteration number and the target function threshold, and solve the adaptive coefficient k of the converter by using the interior point method a The optimal solution in this working condition.

2. The steady-state margin based prediction adaptive hybrid control method for a converter according to claim 1, wherein, In the step S1, the method for constructing the adaptive hybrid synchronous control dq impedance model of the grid-connected VSC converter comprises the following steps: Step S101: building a physical model of the converter with adaptive hybrid synchronous control; Step S102: based on the dynamic process of the phase-locked loop and the power synchronous link, the small-signal dq impedance modeling method is used to construct the small-signal equations of the hybrid synchronous link, the reactive power-voltage droop control link, the voltage control link, the current control link, the coordinate transformation link, and the physical circuit according to the physical model; Step S103: the small-signal equations established in step S102 are solved to obtain the dq impedance transfer function of the converter.

3. The steady-state margin based prediction adaptive hybrid control method for a converter according to claim 1, wherein, In the step S2, the method for constructing the steady-state margin prediction model based on the DNN comprises the following steps: Step S201: analyze the key factors affecting the small signal stability of the power grid-converter system, screen the key features affecting the small signal stability of the power grid-converter system, and use feature engineering to construct the input label X of the steady-state margin prediction model I ; Step S202: considering the number of input labels and output labels, determining the DNN structure of the steady-state margin prediction model, and selecting a loss function L for optimizing the weights and biases of the steady-state margin prediction model in step S3; Step S203: The steady-state margin obtained in step S1 is used as the output label Y I , combined with the input label X in S201 I Construct sample data set D I ={X I , Y I }, used for regression training of the steady-state margin prediction model in step S3.

4. The steady-state margin based prediction adaptive hybrid control method for a converter according to claim 3, wherein, The input label X of the steady state margin prediction model in the step S201 I is represented as: ； wherein k aI is the converter adaptive coefficient in the sample data, P refI is the converter operating condition characteristic active power reference value in the sample data, SCR I is the grid side short circuit ratio in the sample data.

5. The steady-state margin based prediction adaptive hybrid control method for a converter according to claim 3, wherein, The DNN structure of the steady-state margin prediction model in step S202 is a four-layer structure, which includes an input layer, an output layer, and two hidden layers. The activation function type of the hidden layer neurons is selected as the Sigmoid function, and the activation function of the output layer neurons is a linear function. The output of each layer of neurons is represented by the following formula: ； ； where w and b represent weights and biases, respectively; n represents the layer number of the neuron, and the input layer is defined as the 0th layer; p represents the pth neuron in the n-1th layer; q represents the qth neuron in the nth layer; k represents the number of neurons in the n-1th layer; m represents the number of neurons in the nth layer; h q represents the output of the qth neuron in the n hidden layer; w p,q represents the weight between the pth neuron in the n-1th layer and the qth neuron in the nth layer; x p is the output of the pth neuron in the n-1th layer; b q is the bias between the bias node in the n-1th layer and the qth neuron in the nth layer; f(X I ) represents the output of the output layer neuron.

6. The steady-state margin based prediction adaptive hybrid control method for a converter as claimed in claim 3, wherein, The loss function L in step S202 is represented as: ； where MSE is the mean square error, N represents the number of samples, I represents the Ith sample data, Y I is the steady state margin solved in step S1, f(X I ) represents the output of the output layer neuron.

7. The steady-state margin based prediction adaptive hybrid control method for a converter according to claim 3, wherein, The output label Y in the step S203 I is represented as: ； where GM I is the steady state margin in the sample data.

8. The steady-state margin based prediction adaptive hybrid control method for a converter as claimed in claim 1, wherein, The training data in step S3 includes the short-circuit ratio of the grid side, the reference value of the active power, the PCC point voltage dq-axis component, the PCC point current dq-axis component, the converter bridge arm voltage dq-axis component, and the converter bridge arm current dq-axis component.

Images