A network-constructing inverter control method based on physical information neural network

Abstract

The application discloses a network-constructing inverter control method based on a physical information neural network, and the method comprises the following steps: constructing an initial pulse width modulation parameter prediction model; training the initial pulse width modulation parameter prediction model according to historical monitoring data, optimizing model parameters through a total loss function, and determining a preset pulse width modulation parameter prediction model; collecting monitoring data; constructing an input vector x according to the monitoring data; inputting the input vector x into the preset pulse width modulation parameter prediction model to output pulse width modulation parameters; and controlling and adjusting the network-constructing inverter based on the pulse width modulation parameters. The application can automatically calculate optimal signal output based on actual working conditions, and can effectively improve the frequency response performance, fault ride-through performance and power grid strength adaptability of the system.

Description

Technical Field

[0001] This application relates to the field of power system grid-type inverter control technology, and more specifically, to a grid-type inverter control method based on physical information neural networks. Background Technology

[0002] Building a clean, low-carbon, safe, and efficient modern energy system and a new type of power system based on new energy sources, namely a "dual-high" power system (high proportion of new energy and high proportion of power electronic equipment), has become an important trend in power system development.

[0003] Compared to traditional grid-following (GFL) grid connection technology, grid-building (GFM) grid connection technology possesses autonomous control capabilities, actively establishing grid frequency and voltage. This significantly improves system frequency regulation and dynamic damping performance, making it a key technology for future renewable energy-dominated power system grid connection. However, current grid-building technologies are designed based on simplified theoretical models and nominal parameters. Considering factors such as grid environment, uncertainties in grid impedance, equipment parameter drift, nonlinear saturation, and multi-machine interaction, actual engineering conditions often differ from theoretical designs. This results in the pulse width modulation reference voltage signal output by the control system often approaching optimality only under a specific design condition, significantly deviating from the optimal control strategy under other conditions.

[0004] In summary, there is an urgent need to explore a grid-connected inverter control method based on Physical Information Neural Network (PINN) to solve the problem that the pulse width modulation reference voltage signal output by the control system is not the optimal signal due to the gap between actual operating conditions and theoretical design, and to effectively improve the frequency response performance, fault ride-through performance and grid strength adaptability of the system. Summary of the Invention

[0005] In view of the above problems, the purpose of this invention is to provide a grid-type inverter control method based on physical information neural network. The method inputs the monitoring data obtained from the actual operating conditions into a preset pulse width modulation parameter prediction model constructed based on physical information neural network to calculate the optimal signal output, which can effectively improve the frequency response performance, fault ride-through performance and grid strength adaptability of the system.

[0006] The first aspect of this invention provides a control method for a network-type inverter based on a physical information neural network, comprising:

[0007] Construct an initial pulse width modulation parameter prediction model;

[0008] The initial pulse width modulation parameter prediction model is trained based on historical monitoring data. During the training process, data loss function, physical loss function, current limiting loss function, frequency change rate loss function, and voltage change rate loss function are calculated respectively, and the total loss function is determined by combining the corresponding loss function weights. Among them, the physical loss function is obtained by calculating the constraint residuals of each influencing parameter and weighting the calculation results. The historical monitoring data includes at least the three-phase current, DC bus voltage, grid connection point three-phase voltage, DC capacitance, frequency change rate, and voltage change rate obtained during the historical monitoring process.

[0009] The model parameters are optimized using the total loss function to determine the prediction model for the preset pulse width modulation parameters.

[0010] Collect monitoring data;

[0011] Construct an input vector x based on the monitoring data;

[0012] The input vector x is input into a preset pulse width modulation parameter prediction model, and the pulse width modulation parameters are output.

[0013] The grid-type inverter is controlled and adjusted based on the pulse width modulation parameters.

[0014] In this scheme, the preset pulse width modulation parameter prediction model includes an input layer, a PINN hidden layer and an output layer. The PINN hidden layer consists of M hidden layers, each of which is a fully connected structure. Each hidden layer contains 128 neurons and is activated by the Sigmoid activation function.

[0015] In this scheme, the calculation expression of the preset pulse width modulation parameter prediction model is specifically as follows:

[0016] ;

[0017] Among them, h (0) Indicates input layer activation, h (m) W represents the activation output of the m-th hidden layer, where M represents the total number of hidden layers. (m) Let b represent the weight matrix of the m-th layer. (m) Let h represent the bias vector of the m-th layer, σ represent the sigmoid activation function, and h represent the bias vector of the m-th layer. (m+1) This indicates the activation of the (m+1)th hidden layer. W represents the output layer result. M h represents the output layer weight matrix. M b represents the input of the output layer. M This represents the output layer bias matrix.

[0018] This plan also includes:

[0019] The monitoring data includes the three-phase current i output by the grid-side converter. Labc DC bus voltage u dc and the three-phase voltage at the grid connection point v gabc ;

[0020] The pulse width modulation parameters include the phase angle θ required for pulse width modulation. g d-axis reference voltage signal u td * and q-axis reference voltage signal u tq *

[0021] In this scheme, the expression for the input vector x is specifically as follows:

[0022] ;

[0023] Among them, C dc This represents DC capacitance, and ROCOF and ROCOV represent the rate of change of frequency and the rate of change of voltage, respectively.

[0024] In this scheme, the physical loss function is calculated as follows:

[0025] ;

[0026] Among them, L Physics Let r represent the physical loss function. ω Represents the frequency dynamic residual, r Ld and r Lq Represent the dynamic residuals of the AC side d-axis inductance and q-axis inductance, respectively. dc Represents the DC-side power balance residual, r θ The phase angle generates the constraint residual, r P The residuals representing the active power calculation constraints, r Q The residuals representing the reactive power calculation constraints, r vdq Represents the voltage coordinate transformation constraint residual, r idq λ represents the constraint residual of the current coordinate transformation. ω λ represents the weighting coefficient of the frequency dynamics residual. i λ represents the weighting coefficient of the dynamic residual of the AC-side inductor. dc λ represents the weighting coefficient for the DC-side power balance residual. θ λ represents the phase angle generation constraint residual weighting coefficient. PQ λ represents the weighting coefficient of the power calculation constraint residual. dq This represents the weighting coefficient of the coordinate transformation constraint residual.

[0027] In this solution, the calculation method of the data loss function is as follows:

[0028] ;

[0029] Among them, L data Represents the data loss function; N data Indicates the number of training samples; Indicates the prediction result; y j Indicates the actual result;

[0030] The specific method for calculating the current limiting loss function is as follows:

[0031] ;

[0032] Among them, L current N represents the current limiting loss function. current I represents the number of training samples. total (x j ) represents the current amplitude of the j-th sample, I peak This indicates the inverter's rated maximum peak current;

[0033] The specific method for calculating the frequency change rate loss function is as follows:

[0034] ;

[0035] Among them, L ROCOF N represents the loss function based on the rate of change of frequency. ROCOF X represents the number of training samples. jR Indicates in sample x j The rate of change of frequency at that location;

[0036] The specific method for calculating the voltage change rate loss function is as follows:

[0037] ;

[0038] Among them, L ROCOV N represents the voltage change rate loss function. ROCOV X represents the number of training samples. jV Indicates in sample x j The rate of change of voltage at that point.

[0039] In this scheme, the specific method for calculating the frequency dynamic residual is as follows:

[0040] ;

[0041] Where, r ω N represents the frequency dynamic residual. ω The number of training samples is represented by ω and ω0, respectively, which represent the angular velocity and rated angular velocity of the grid-side converter. J represents the moment of inertia, D represents the damping coefficient, and P represents the damping coefficient. jin and Pjout These represent the input power and output power of the j-th sample grid-side converter, respectively.

[0042] The specific method for calculating the dynamic residual of the AC side inductor is as follows:

[0043] ;

[0044] Where, r Ld and r Lq N represents the dynamic residual of the AC side d-axis inductance and the dynamic residual of the AC side q-axis inductance, respectively. Ld and N Lq Both represent the number of training samples, i Ld and i Lq L represents the d-axis current and q-axis current of the grid-side converter, respectively. f U represents the equivalent inductance of the grid-side converter. jLd and u jLq The j-th sample represents the d-axis voltage and q-axis voltage of the grid-side converter, respectively. jgd and u jgq Let represent the d-axis voltage and q-axis voltage of the j-th sample grid connection point, respectively.

[0045] In this scheme, the calculation method for the DC-side power balance residual is as follows:

[0046] ;

[0047] Where, r dc N represents the DC-side power balance residual. dc u represents the number of training samples. dc C represents the DC bus voltage. dc Indicates DC capacitance, i jdc,in P represents the input current of the grid-side converter for the j-th sample. j U represents the active power of the j-th sample. jdc This represents the DC voltage of the j-th sample;

[0048] The specific method for calculating the phase angle generation constraint residual is as follows:

[0049] ;

[0050] Where, r θ N represents the phase angle generating constraint residual. θ θ represents the number of training samples. jg Let ω0 represent the phase angle of the j-th sample, and C represent the rated angular velocity of the grid-side converter. dc For DC capacitor, u jdc U represents the DC voltage of the j-th sample. jdcrefH represents the DC voltage reference value of the j-th sample. v D is the virtual inertia coefficient. v θ is the virtual damping coefficient. g Let represent the phase angle of the grid-side converter, and s be the Laplace operator.

[0051] In this scheme, the specific method for calculating the power calculation constraint residual is as follows:

[0052] ;

[0053] Where, r P and r Q N represents the residuals of active power calculation constraints and reactive power calculation constraints, respectively. P and N Q Both represent the number of training samples, P j and Q j Let v represent the active power and reactive power of the j-th sample, respectively. jgd and v jgq Let i represent the d-axis voltage and q-axis voltage at the grid connection point of the j-th sample, respectively. jLd and i jLq These represent the d-axis output current and q-axis output current of the j-th sample of the grid-side converter, respectively.

[0054] The specific method for calculating the coordinate transformation constraint residual is as follows:

[0055] ;

[0056] Where, r vdq and r idq N represents the voltage coordinate transformation constraint residual and the current coordinate transformation constraint residual, respectively. vdq and N idq Both represent the number of training samples, θ jg V represents the phase angle of the j-th sample. jgd and v jgq Let v represent the d-axis voltage and q-axis voltage of the j-th sample grid connection point, respectively. jga v jgb and v jgc Let i represent the phase a voltage, phase b voltage, and phase c voltage at the j-th sample grid connection point, respectively. jLd and i jLq Let i represent the d-axis output current and q-axis output current of the j-th sample of the grid-side converter, respectively. jLa i jLb and i jLc These represent the phase a, phase b, and phase c output currents of the j-th sample of the grid-side converter, respectively.

[0057] This invention discloses a control method for a grid-type inverter based on a physical information neural network. The method includes: constructing an initial pulse width modulation (PWM) parameter prediction model; training the initial PWM parameter prediction model based on historical monitoring data, optimizing the model parameters through a total loss function, and determining a preset PWM parameter prediction model; collecting monitoring data; constructing an input vector x based on the monitoring data; inputting the input vector x into the preset PWM parameter prediction model and outputting PWM parameters; and controlling and adjusting the grid-type inverter based on the PWM parameters.

[0058] The reference voltage signal u required by the pulse width modulation of this invention td *、u tq * and phase angle θ g Without the need for traditional control procedures, the signal is automatically generated directly by the PINN module. Traditional control procedures are based on simplified theoretical models and nominal parameter designs, and actual operating conditions often differ from theoretical designs, resulting in the pulse width modulation reference voltage signal output by the control system not being the optimal signal. In contrast, the physical information neural network module automatically calculates the optimal signal output based on actual operating conditions, which can effectively improve the system's frequency response performance, fault ride-through performance, and grid strength adaptability. Attached Figure Description

[0059] Figure 1 A flowchart of a network inverter control method based on a physical information neural network provided by the present invention is shown;

[0060] Figure 2 The diagram shows the overall flow of the network inverter control method based on physical information neural network provided by the present invention. Detailed Implementation

[0061] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0062] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0063] Figure 1 A flowchart of a network inverter control method based on a physical information neural network provided by the present invention is shown.

[0064] like Figure 1As shown, this invention discloses a control method for a network-type inverter based on a physical information neural network, comprising:

[0065] S101, Construct the initial pulse width modulation parameter prediction model;

[0066] S102, The initial pulse width modulation parameter prediction model is trained based on historical monitoring data. During the training process, the data loss function, physical loss function, current limiting loss function, frequency change rate loss function, and voltage change rate loss function are calculated respectively. The total loss function is determined by combining the corresponding loss function weights. Among them, the physical loss function is obtained by calculating the constraint residuals of each influencing parameter and weighting the calculation results.

[0067] S103, optimize the model parameters through the total loss function to determine the prediction model with preset pulse width modulation parameters;

[0068] S104, Collect monitoring data;

[0069] S105, Construct the input vector x based on the monitoring data;

[0070] S106, Input the input vector x into the preset pulse width modulation parameter prediction model, and output the pulse width modulation parameters;

[0071] S107 controls and adjusts the grid-type inverter based on pulse width modulation parameters.

[0072] According to embodiments of the present invention, such as Figure 2 As shown, the entire control system includes the renewable energy power plant, the generator-side converter (MSC), the grid-side converter (GSC), and the power grid. The electricity generated by the renewable energy power plant is rectified by the generator-side converter, then inverted by the grid-side converter before being fed into the power grid. Figure 2 In the middle, P in and P out These represent the input active power and output active power of the GSC, respectively. dc C represents the DC bus voltage. dc This represents the DC capacitance, ROCOF and ROCOV represent the rate of change of frequency and the rate of change of voltage, respectively. Labc L represents the output current of the GSC. f This represents the equivalent inductance of GSC, v gabc Indicates the three-phase voltage at the grid connection point, s gabc This indicates the trigger pulse for the GSC.

[0073] Monitoring data is collected at preset time intervals or in real time. The monitoring data includes at least the three-phase current i output by the grid-side converter. Labc DC bus voltage u dc and the three-phase voltage at the grid connection point vgabc Based on the obtained monitoring data, an input vector x is constructed and fed into the trained PINN model (i.e., a pre-defined pulse width modulation parameter prediction model). After model analysis and calculation, the output includes the phase angle θ required for pulse width modulation. g d-axis reference voltage signal u td * and q-axis reference voltage signal u tq The pulse width modulation parameters are precisely controlled to adjust the grid-connected inverter. Based on monitoring data under actual operating conditions, the optimal signal output is automatically calculated, effectively improving the system's frequency response performance, fault ride-through performance, and grid strength adaptability.

[0074] The training process for the initial pulse width modulation parameter prediction model is based on the PyTorch framework and uses the Adam optimization algorithm. All training data is loaded and converted into PyTorch tensors. The training data mainly comes from the actual data of the power grid company, i.e., historical monitoring data, or simulation data generated through a large number of simulations.

[0075] The initial (preset) pulse width modulation parameter prediction model employs a fully connected feedforward neural network, comprising an input layer, a PINN hidden layer, and an output layer. Historical monitoring data includes the three-phase current i output from the grid-side converter. Labc (A-phase output current, B-phase output current, and C-phase output current), current amplitude, DC bus voltage u dc Three-phase voltage at grid connection point V gabc (Including phase A voltage, phase B voltage, and phase C voltage), DC capacitance, rate of change of frequency, rate of change of voltage, equivalent inductance, DC voltage, d-axis and q-axis currents, d-axis and q-axis voltages of the grid-side converter, d-axis and q-axis voltages at the grid connection point, angular velocity and phase angle of the grid-side transmission, input current of the grid-side converter, active power and reactive power, and virtual inertia coefficient and virtual damping coefficient of the motor, etc. Each parameter can be obtained through corresponding monitoring or sensor equipment, or calculated using existing formulas from the collected parameters.

[0076] The specific method for calculating the total loss function is as follows:

[0077] ;

[0078] Where L represents the total loss function, L data Represents the data loss function, L Physics Represents the physical loss function, L current Represents the current limiting loss function, L ROCOF Represents the loss function of frequency change rate, L ROCOV Represents the voltage change rate loss function, λPDE Represents the weights of the physical loss function, λ current Represents the weights of the current limiting loss function, λ ROCOF The weights of the frequency change rate loss function, λ ROCOV This represents the weight of the voltage change rate loss function.

[0079] Wherein, the physical loss function weight λ PDE Weight λ of the current limiting loss function current , frequency change rate loss function weight λ ROCOF and the weight λ of the voltage change rate loss function ROCOV The specific values ​​are set by those skilled in the art based on actual needs.

[0080] Among them, the constraint residuals of each influencing parameter include the frequency dynamic residual r ω AC side d-axis inductance dynamic residual r Ld and q-axis inductance dynamic residual r Lq DC-side power balance residual r dc Phase angle generation constraint residual r θ Active power calculation constraint residual r P Reactive power calculation constraint residual r Q Voltage coordinate transformation constraint residual r vdq and the current coordinate transformation constraint residual r idq The physical loss function is determined by multiplying the constraint residuals of each influencing parameter by their corresponding residual weighting coefficients and summing the calculation results.

[0081] The forward propagation algorithm calculates the predicted output value based on the current parameters and compares it with the actual value. The total loss is then calculated using the total loss function formula to determine the difference between the predicted and actual values. Finally, the backpropagation algorithm calculates the gradient of the loss function with respect to the network parameters, thus determining the direction of parameter optimization. The gradient expression is:

[0082] ;

[0083] in, Let L represent the gradient with respect to θ, L represent the total loss, and θ represent all trainable parameters of the model.

[0084] Furthermore, the Adam algorithm is used to update the weights and biases of each layer. Using the gradients just calculated, the network parameters are modified to make the predicted output value of the next round of network computation closer to the actual value. The expression is as follows:

[0085] ;

[0086] Where, θ t θ represents the parameters after the t-th iteration;t-1 The parameters represent the values ​​at the end of the (t-1)th iteration; α represents the learning rate. This represents the first moment estimate of the gradient; This represents the second moment estimate of the gradient; Represents a very small positive number (such as 10). -8 Or 10 -7 ), to prevent the denominator from being 0.

[0087] Furthermore, the training steps are iterated by repeating the process of "forward propagation → loss calculation → back propagation → parameter update". During the training process, the total loss gradually decreases until convergence (e.g., 7000 iterations, or the difference between the predicted value and the actual value meets the system's preset requirements). Finally, a preset pulse width modulation parameter prediction model that can be used for real-time control is obtained. The preset pulse width modulation parameter prediction model can replace the traditional control structure (e.g., the dual closed-loop voltage-current control structure).

[0088] According to an embodiment of the present invention, it further includes:

[0089] The monitoring data includes the three-phase current i output from the grid-side converter. Labc DC bus voltage u dc and the three-phase voltage at the grid connection point v gabc ;

[0090] Pulse width modulation parameters include the phase angle θ required for pulse width modulation. g d-axis reference voltage signal u td * and q-axis reference voltage signal u tq *

[0091] It should be noted that the three-phase current i output by the grid-side converter Labc The DC bus voltage u can be obtained through current sensors such as Hall effect sensors. dc The three-phase voltage at the grid connection point can be obtained through voltage sensors such as isolation voltage sensors. gabc Voltage can be obtained through voltage sensors such as voltage transformers.

[0092] According to an embodiment of the present invention, the expression for the input vector x is specifically as follows:

[0093] ;

[0094] Among them, C dc This represents DC capacitance, and ROCOF and ROCOV represent the rate of change of frequency and the rate of change of voltage, respectively.

[0095] It should be noted that the DC capacitor C dc This information can be found on the capacitor's nameplate or in the datasheet provided with the capacitor.

[0096] ROCOF (Rate of Change of Freguency) represents the rate of change of frequency and is a key indicator for measuring the instantaneous rate of change of frequency in a power system. Its unit is usually Hz / s.

[0097] ROCOV (Rate of Change of Voltage) represents the rate of change of voltage, describing the instantaneous rate of change of voltage, and its unit is usually V / s.

[0098] According to an embodiment of the present invention, the preset pulse width modulation parameter prediction model adopts a fully connected feedforward neural network, including an input layer, a PINN hidden layer, and an output layer. The input layer is an input vector x containing 12-dimensional features. The PINN hidden layer consists of M hidden layers, each of which is a fully connected structure, containing 128 neurons, and each hidden layer is activated using the Sigmoid activation function. The output layer is a linear layer, without activation, directly outputting the reference voltage signal and phase angle required for pulse width modulation.

[0099] It should be noted that, except for the last layer (output layer) which remains linear, the rest are forward propagation structures, that is, each hidden layer first undergoes a linear transformation and then is activated by the Sigmoid activation function.

[0100] The expression for the sigmoid activation function is:

[0101] ;

[0102] Where σ represents the sigmoid activation function; x represents the input vector.

[0103] In addition, the number of hidden layers is usually set to three, i.e., M=3. Those skilled in the art can dynamically adjust the number of hidden layers according to actual needs.

[0104] According to an embodiment of the present invention, the calculation expression of the preset pulse width modulation parameter prediction model is specifically as follows:

[0105] ;

[0106] Among them, h (0) Indicates input layer activation, h (m) W represents the activation output of the m-th hidden layer, where M represents the total number of hidden layers. (m) Let b represent the weight matrix of the m-th layer. (m) Let h represent the bias vector of the m-th layer, σ represent the sigmoid activation function, and h represent the bias vector of the m-th layer. (m+1) This indicates the activation of the (m+1)th hidden layer. W represents the output layer result. M h represents the output layer weight matrix.M b represents the input of the output layer. M This represents the output layer bias matrix.

[0107] It should be noted that the input vector x is fed into the preset pulse width modulation parameter prediction model, activating the model's input layer. The input layer receives the input vector x and passes it to the PINN hidden layer. The PINN hidden layer performs feature extraction and processing on the input vector x, and the output layer converts the analysis and processing of the PINN hidden layer into the final output, obtaining the phase angle θ required for pulse width modulation. g d-axis reference voltage signal u td * and q-axis reference voltage signal u tq * Pulse width modulation parameters.

[0108] The output expression is:

[0109] ;

[0110] in, This represents the output of the prediction model with preset pulse width modulation parameters.

[0111] According to an embodiment of the present invention, the method for calculating the physical loss function is as follows:

[0112] ;

[0113] Among them, L Physics Let r represent the physical loss function. ω Represents the frequency dynamic residual, r Ld and r Lq Represent the dynamic residuals of the AC side d-axis inductance and q-axis inductance, respectively. dc Represents the DC-side power balance residual, r θ The phase angle generates the constraint residual, r P The residuals representing the active power calculation constraints, r Q The residuals representing the reactive power calculation constraints, r vdq Represents the voltage coordinate transformation constraint residual, r idq λ represents the constraint residual of the current coordinate transformation. ω λ represents the weighting coefficient of the frequency dynamics residual. i λ represents the weighting coefficient of the dynamic residual of the AC-side inductor. dc λ represents the weighting coefficient for the DC-side power balance residual. θ λ represents the phase angle generation constraint residual weighting coefficient. PQ λ represents the weighting coefficient of the power calculation constraint residual. dq This represents the weighting coefficient of the coordinate transformation constraint residual.

[0114] It should be noted that the frequency dynamic residual weighting coefficient λ ω AC side inductor dynamic residual weighting coefficient λ i DC-side power balance residual weighting coefficient λ dc Phase angle generation constraint residual weighting coefficient λ θ Power calculation constraint residual weighting coefficient λ PQ And coordinate transformation constraint residual weight coefficient λ dq All settings are configured by those skilled in the art based on actual needs. Among them, λ ω +2λ i +λ dc +λ θ +2λ PQ +2λ dq =1.

[0115] According to an embodiment of the present invention, the method for calculating the data loss function is as follows:

[0116] ;

[0117] Among them, L data Represents the data loss function; N data Indicates the number of training samples; This represents the prediction result of the j-th sample after inputting the preset pulse width modulation parameters into the prediction model; y j This represents the actual result corresponding to the j-th sample;

[0118] The specific method for calculating the current limiting loss function is as follows:

[0119] ;

[0120] Among them, L current N represents the current limiting loss function. current I represents the number of training samples. total (x j ) represents the current amplitude of the j-th sample, I peak This indicates the inverter's rated maximum peak current (which can be determined from the inverter's factory parameters).

[0121] The specific method for calculating the frequency change rate loss function is as follows:

[0122] ;

[0123] Among them, L ROCOF N represents the loss function based on the rate of change of frequency. ROCOF X represents the number of training samples. jR Indicates in sample x j The rate of change of frequency at that location;

[0124] The specific method for calculating the voltage change rate loss function is as follows:

[0125] ;

[0126] Among them, L ROCOV N represents the voltage change rate loss function. ROCOV X represents the number of training samples. jV Indicates in sample x j The rate of change of voltage at that point.

[0127] According to an embodiment of the present invention, the method for calculating the frequency dynamic residual is as follows:

[0128] ;

[0129] Where, r ω N represents the frequency dynamic residual. ω The number of training samples is represented by ω and ω0, which represent the angular velocity and rated angular velocity of the grid-side converter (pre-set by the system), respectively. J represents the moment of inertia, D represents the damping coefficient, and P represents the damping coefficient. jin and P jout These represent the input power and output power of the j-th sample grid-side converter, respectively.

[0130] The specific method for calculating the dynamic residual of the AC side inductance is as follows:

[0131] ;

[0132] Where, r Ld and r Lq N represents the dynamic residual of the AC side d-axis inductance and the dynamic residual of the AC side q-axis inductance, respectively. Ld and N Lq Both represent the number of training samples, i Ld and i Lq L represents the d-axis current and q-axis current of the grid-side converter, respectively. f U represents the equivalent inductance of the grid-side converter. jLd and u jLq The j-th sample represents the d-axis voltage and q-axis voltage of the grid-side converter, respectively. jgd and u jgq Let represent the d-axis voltage and q-axis voltage of the j-th sample grid connection point, respectively.

[0133] According to an embodiment of the present invention, the method for calculating the DC-side power balance residual is as follows:

[0134] ;

[0135] Where, r dc N represents the DC-side power balance residual. dcu represents the number of training samples. dc C represents the DC bus voltage. dc Indicates DC capacitance, i jdc,in P represents the input current of the grid-side converter for the j-th sample. j U represents the active power of the j-th sample. jdc This represents the DC voltage of the j-th sample;

[0136] The specific method for calculating the phase angle generation constraint residual is as follows:

[0137] ;

[0138] Where, r θ N represents the phase angle generating constraint residual. θ θ represents the number of training samples. jg Let ω0 represent the phase angle of the j-th sample, and C represent the rated angular velocity of the grid-side converter. dc For DC capacitor, u jdc U represents the DC voltage of the j-th sample. jdcref H represents the DC voltage reference value of the j-th sample (pre-set by the system). v D is the virtual inertia coefficient. v θ is the virtual damping coefficient. g Let represent the phase angle of the grid-side converter, and s be the Laplace operator.

[0139] Virtual inertia coefficient H v and virtual damping coefficient D v The expressions are as follows:

[0140] ;

[0141] Where H represents the inertial time constant, D represents the damping coefficient, S0 represents the rated apparent power of the system, and ω0 represents the rated angular velocity.

[0142] The inertial time constant H is obtained by converting the moment of inertia J, and its expression is:

[0143] .

[0144] According to an embodiment of the present invention, the method for calculating the power calculation constraint residual is as follows:

[0145] ;

[0146] Where, r P and r Q N represents the residuals of active power calculation constraints and reactive power calculation constraints, respectively. P and N QBoth represent the number of training samples, P j and Q j Let v represent the active power and reactive power of the j-th sample, respectively. jgd and v jgq Let i represent the d-axis voltage and q-axis voltage at the grid connection point of the j-th sample, respectively. jLd and i jLq These represent the d-axis output current and q-axis output current of the j-th sample of the grid-side converter, respectively.

[0147] The specific method for calculating coordinate transformation constraint residuals is as follows:

[0148] ;

[0149] Where, r vdq and r idq N represents the voltage coordinate transformation constraint residual and the current coordinate transformation constraint residual, respectively. vdq and N idq Both represent the number of training samples, θ jg V represents the phase angle of the j-th sample. jgd and v jgq Let v represent the d-axis voltage and q-axis voltage of the j-th sample grid connection point, respectively. jga v jgb and v jgc Let i represent the phase a voltage, phase b voltage, and phase c voltage at the j-th sample grid connection point, respectively. jLd and i jLq Let i represent the d-axis output current and q-axis output current of the j-th sample of the grid-side converter, respectively. jLa i jLb and i jLc These represent the phase a, phase b, and phase c output currents of the j-th sample of the grid-side converter, respectively.

[0150] Wherein, the number of training samples N data N current N ROCOF N ROCOV N ω N Ld N Lq N dc N θ N P N Q N vdq and N idq The number of samples is the same for all of them, which is the total number of training data. The different letters representing the number of training samples are only used to distinguish the calculation of different loss functions or constraint residuals.

[0151] All information (including but not limited to user equipment information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals (including but not limited to signals transmitted between user terminals and other devices) involved in this application have been authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the "historical monitoring data" and "monitoring data" mentioned in this disclosure were obtained under full authorization.

[0152] This invention discloses a control method for a grid-connected inverter based on a physical information neural network. The method includes: constructing an initial pulse width modulation (PWM) parameter prediction model; training the initial PWM parameter prediction model based on historical monitoring data, optimizing the model parameters using a total loss function, and determining a preset PWM parameter prediction model; collecting monitoring data; constructing an input vector x based on the monitoring data; inputting the input vector x into the preset PWM parameter prediction model to output PWM parameters; and controlling and adjusting the grid-connected inverter based on the PWM parameters. This invention automatically calculates the optimal signal output based on actual operating conditions, which can effectively improve the system's frequency response performance, fault ride-through performance, and grid strength adaptability.

[0153] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0154] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0155] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0156] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0157] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

Claims

1. A control method for a network-type inverter based on a physical information neural network, characterized in that, include: Construct an initial pulse width modulation parameter prediction model; The initial pulse width modulation parameter prediction model is trained based on historical monitoring data. During the training process, data loss function, physical loss function, current limiting loss function, frequency change rate loss function, and voltage change rate loss function are calculated respectively, and the total loss function is determined by combining the corresponding loss function weights. Among them, the physical loss function is obtained by calculating the constraint residuals of each influencing parameter and weighting the calculation results. The historical monitoring data includes at least the three-phase current, DC bus voltage, grid connection point three-phase voltage, DC capacitance, frequency change rate, and voltage change rate obtained during the historical monitoring process. The model parameters are optimized using the total loss function to determine the prediction model for the preset pulse width modulation parameters. Collect monitoring data; Construct an input vector x based on the monitoring data; The input vector x is input into a preset pulse width modulation parameter prediction model, and the pulse width modulation parameters are output. The grid-type inverter is controlled and adjusted based on the pulse width modulation parameters. The calculation expression for the preset pulse width modulation parameter prediction model is as follows: ； Among them, h (0) Indicates input layer activation, h (m) W represents the activation output of the m-th hidden layer, where M represents the total number of hidden layers. (m) Let b represent the weight matrix of the m-th layer. (m) Let h represent the bias vector of the m-th layer, σ represent the sigmoid activation function, and h represent the bias vector of the m-th layer. (m+1) This indicates the activation of the (m+1)th hidden layer. W represents the output layer result. M h represents the output layer weight matrix. M b represents the input of the output layer. M This represents the output layer bias matrix; The specific expression for the input vector x is: ； Among them, C dc ROCOF and ROCOV represent the rate of change of frequency and the rate of change of voltage, respectively. The specific method for calculating the physical loss function is as follows: ； Among them, L Physics Let r represent the physical loss function. ω Represents the frequency dynamic residual, r Ld and r Lq Represent the dynamic residuals of the AC side d-axis inductance and q-axis inductance, respectively. dc Represents the DC-side power balance residual, r θ The phase angle generates the constraint residual, r P The residuals representing the active power calculation constraints, r Q The residuals representing the reactive power calculation constraints, r vdq Represents the voltage coordinate transformation constraint residual, r idq λ represents the constraint residual of the current coordinate transformation. ω λ represents the weighting coefficient of the frequency dynamics residual. i λ represents the weighting coefficient of the dynamic residual of the AC-side inductor. dc λ represents the weighting coefficient for the DC-side power balance residual. θ λ represents the phase angle generation constraint residual weighting coefficient. PQ λ represents the weighting coefficient of the power calculation constraint residual. dq This represents the weighting coefficient of the coordinate transformation constraint residual.

2. The control method for a network-type inverter based on a physical information neural network according to claim 1, characterized in that, The preset pulse width modulation parameter prediction model includes an input layer, a PINN hidden layer, and an output layer. The PINN hidden layer consists of M hidden layers, each of which is a fully connected structure. Each hidden layer contains 128 neurons and is activated using the Sigmoid activation function.

3. The control method for a network-type inverter based on a physical information neural network according to claim 1, characterized in that, Also includes: The monitoring data includes the three-phase current i output by the grid-side converter. Labc DC bus voltage u dc and the three-phase voltage at the grid connection point v gabc ; The pulse width modulation parameters include the phase angle θ required for pulse width modulation. g d-axis reference voltage signal u td * and q-axis reference voltage signal u tq * 4. The control method for a network-type inverter based on a physical information neural network according to claim 1, characterized in that, The specific method for calculating the data loss function is as follows: ； Among them, L data Represents the data loss function; N data Indicates the number of training samples; Indicates the prediction result; y j Indicates the actual result; The specific method for calculating the current limiting loss function is as follows: ； Among them, L current N represents the current limiting loss function. current I represents the number of training samples. total (x j ) represents the current amplitude of the j-th sample, I peak This indicates the inverter's rated maximum peak current; The specific method for calculating the frequency change rate loss function is as follows: ； Among them, L ROCOF N represents the loss function based on the rate of change of frequency. ROCOF X represents the number of training samples. jR Indicates in sample x j The rate of change of frequency at that location; The specific method for calculating the voltage change rate loss function is as follows: ； Among them, L ROCOV N represents the voltage change rate loss function. ROCOV X represents the number of training samples. jV Indicates in sample x j The rate of change of voltage at that point.

5. The control method for a network-type inverter based on a physical information neural network according to claim 1, characterized in that, The specific method for calculating the frequency dynamic residual is as follows: ； Where, r ω N represents the frequency dynamic residual. ω The number of training samples is represented by ω and ω0, respectively, which represent the angular velocity and rated angular velocity of the grid-side converter. J represents the moment of inertia, D represents the damping coefficient, and P represents the damping coefficient. jin and P jout These represent the input power and output power of the j-th sample grid-side converter, respectively. The specific method for calculating the dynamic residual of the AC side inductor is as follows: ； Where, r Ld and r Lq N represents the dynamic residual of the AC side d-axis inductance and the dynamic residual of the AC side q-axis inductance, respectively. Ld and N Lq Both represent the number of training samples, i Ld and i Lq L represents the d-axis current and q-axis current of the grid-side converter, respectively. f U represents the equivalent inductance of the grid-side converter. jLd and u jLq The j-th sample represents the d-axis voltage and q-axis voltage of the grid-side converter, respectively. jgd and u jgq Let represent the d-axis voltage and q-axis voltage of the j-th sample grid connection point, respectively.

6. The control method for a network-type inverter based on a physical information neural network according to claim 1, characterized in that, The specific method for calculating the DC-side power balance residual is as follows: ； Where, r dc N represents the DC-side power balance residual. dc u represents the number of training samples. dc C represents the DC bus voltage. dc Indicates DC capacitance, i jdc,in P represents the input current of the grid-side converter for the j-th sample. j U represents the active power of the j-th sample. jdc This represents the DC voltage of the j-th sample; The specific method for calculating the phase angle generation constraint residual is as follows: ； Where, r θ N represents the phase angle generating constraint residual. θ θ represents the number of training samples. jg Let ω0 represent the phase angle of the j-th sample, and C represent the rated angular velocity of the grid-side converter. dc For DC capacitor, u jdc U represents the DC voltage of the j-th sample. jdcref H represents the DC voltage reference value of the j-th sample. v D is the virtual inertia coefficient. v θ is the virtual damping coefficient. g Let represent the phase angle of the grid-side converter, and s be the Laplace operator.

7. The control method for a network-type inverter based on a physical information neural network according to claim 1, characterized in that, The specific method for calculating the power calculation constraint residual is as follows: ； Where, r P and r Q N represents the residuals of active power calculation constraints and reactive power calculation constraints, respectively. P and N Q Both represent the number of training samples, P j and Q j Let v represent the active power and reactive power of the j-th sample, respectively. jgd and v jgq Let i represent the d-axis voltage and q-axis voltage at the grid connection point of the j-th sample, respectively. jLd and i jLq These represent the d-axis output current and q-axis output current of the j-th sample of the grid-side converter, respectively. The specific method for calculating the coordinate transformation constraint residual is as follows: ； Where, r vdq and r idq N represents the voltage coordinate transformation constraint residual and the current coordinate transformation constraint residual, respectively. vdq and N idq Both represent the number of training samples, θ jg V represents the phase angle of the j-th sample. jgd and v jgq Let v represent the d-axis voltage and q-axis voltage of the j-th sample grid connection point, respectively. jga v jgb and v jgc Let i represent the phase a voltage, phase b voltage, and phase c voltage at the j-th sample grid connection point, respectively. jLd and i jLq Let i represent the d-axis output current and q-axis output current of the j-th sample of the grid-side converter, respectively. jLa i jLb and i jLc These represent the phase a, phase b, and phase c output currents of the j-th sample of the grid-side converter, respectively.

Images