A power control method of a photovoltaic energy storage inverter

Abstract

The application discloses a power control method of a photovoltaic energy storage inverter, adopts a CEEMDAN algorithm to collect current and voltage data of input and output ends of the photovoltaic energy storage inverter, and adaptively processes data with different signal-to-noise ratios to construct a photovoltaic energy storage inverter data model; a CNN-BiGRU network is used to extract data features and perform deep learning training on the photovoltaic energy storage inverter data model; the data features extracted by the CNN-BiGRU network are used to perform omnidirectional time sequence modeling, the BiGRU neural network is connected with an FCFNN full connection neural network and input data sequences are input, so that control data sequences output by the CNN-BiGRU-FCFNN network model are obtained; the DE-BWO algorithm is used to optimize the CNN-BiGRU-FCFNN network model, a deep learning model based on output control of the photovoltaic energy storage inverter is obtained, the successfully trained deep learning model is implanted into a chip module, and the chip module is used to output control output data sequences; the VSG control module is used to calculate output control signals, and the virtual synchronous power output of the photovoltaic energy storage inverter is realized.

Description

Technical Field

[0001] This invention relates to a power control method for a photovoltaic energy storage inverter. Background Technology

[0002] Considering that the rotational inertia of a synchronous generator rotor is beneficial to stabilizing the power system frequency, a virtual synchronous generator (VSG) control technology based on photovoltaic energy storage inverters has been proposed. The core idea is to simulate the rotor motion equations of a synchronous generator within the control strategy of the photovoltaic energy storage inverter. Currently, the design of photovoltaic energy storage inverters, primarily based on traditional PID control methods, struggles to meet the demands of complex and nonlinear systems, easily leading to problems such as runaway control and overshoot. Traditional control methods are highly sensitive to unknown environmental changes, limiting their effectiveness and hindering further performance improvements. Furthermore, they exhibit low adaptability and robustness in identifying and controlling multi-factor interactions.

[0003] In recent years, with the increasingly widespread application of artificial intelligence technology, methods that combine optimization techniques with fused neural networks to achieve Deep Reinforcement Learning (DRL) have attracted much attention. Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) is a signal analysis method based on Empirical Mode Decomposition (EMD). It extracts features from the signal by decomposing the original signal into multiple Intrinsic Mode Functions (IMFs). Compared to traditional EMD, CEEMDAN uses noise to improve its stability and robustness. In terms of neural networks, Convolutional Neural Networks (CNNs) are a widely used deep learning algorithm, capable of directly extracting features from raw data. They have shown good performance in image, speech, and natural language processing, and have also contributed to the development of fusion with other algorithms. Bidirectional Gated Recurrent Unit (BiGRU), evolved from Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) networks, can utilize both preceding and subsequent information for prediction and is increasingly widely used. Beluga whale optimization (BWO), proposed in 2022, is a novel population-based metaheuristic algorithm inspired by the swimming, whaling, and falling behaviors of beluga whales. It possesses certain advantages and potential for solving complex optimization problems.

[0004] Therefore, this invention proposes a power control method for photovoltaic energy storage inverters, which controls the power output of photovoltaic energy storage inverters in real time to achieve more accurate and stable power control. Summary of the Invention

[0005] The present invention provides a power control method for a photovoltaic energy storage inverter in order to solve the problems existing in the prior art.

[0006] The technical solutions adopted in this invention are as follows:

[0007] A power control method for a photovoltaic energy storage inverter, including

[0008] Step 1: The CEEMDAN algorithm is used to collect current and voltage data at the input and output terminals of the photovoltaic energy storage inverter, and the CEEMDAN algorithm is used to adaptively process data with different signal-to-noise ratios in order to construct a data model of the photovoltaic energy storage inverter.

[0009] Step 2: Use the CNN-BiGRU network to extract data features and train deep learning for the photovoltaic energy storage inverter data model;

[0010] Step 3: Based on the BiGRU neural network, perform omnidirectional time series modeling on the data features extracted by the CNN-BiGRU network, connect the output of the BiGRU neural network with the FCFNN fully connected neural network and input the data sequence to obtain the control data sequence output by the CNN-BiGRU-FCFNN network model;

[0011] Step 4: Optimize the parameters of the CNN-BiGRU-FCFNN network model using the DE-BWO algorithm based on differential evolution, and finally complete the learning and training of the CNN-BiGRU-FCFNN network model to obtain a deep learning model based on the output control of photovoltaic energy storage inverter.

[0012] Step 5: Implant the successfully trained deep learning model into the chip module (i.e., GPU + ARM processor), and output the control output data sequence described in Step 3 through the chip module. After the VSG control module calculates and outputs the control signal, the virtual synchronous power output of the photovoltaic energy storage inverter is controlled.

[0013] Furthermore, step one specifically includes the following steps:

[0014] 1) Acquisition and processing of voltage and current data at the input and output terminals of the photovoltaic energy storage inverter;

[0015] 2) The CEEMDAN algorithm is used to process the voltage and current data in step 1) to improve the signal storage accuracy and integrity;

[0016] 3) The permutation entropy method is used to simplify the data model of photovoltaic energy storage inverter, reduce the computational scale, and improve learning efficiency.

[0017] Furthermore, in step two, the upper layer of the CNN-BiGRU network is a CNN network, which is used to accept various decomposed variables that affect the virtual synchronous power generation of the photovoltaic energy storage inverter;

[0018] The lower layer is a BiGRU network, which is used to receive the data feature sequence after the CNN network is processed and to train the data model of the photovoltaic energy storage inverter.

[0019] Furthermore, step four specifically includes the following steps:

[0020] 1) Determine the initial parameters of the differential evolution-based improved BWO algorithm, i.e., the DE-BWO algorithm;

[0021] Maximum number of iterations to initialize the DE-BWO algorithm =1000, population size =100, Leader of the White Whales Beluga whale followers dimensionality This represents the total number of parameters and weight matrices that need to be optimized in the CNN-BiGRU network.

[0022] 2) Initialize the population position for the DE-BWO algorithm The initial positions of each beluga whale are randomly generated within the parameter setting range, and the population fitness value of the DE-BWO algorithm is determined. As shown in the following formula:

[0023] ,

[0024] in, The value function of the CNN-BiGRU-FCFNN network parameters. The target value output by the CNN-BiGRU-FCFNN network model;

[0025] These are the weights of the convolutional layers in a CNN network.

[0026] Calculate the weight matrix for the neurons of the FCFNN fully connected neural network;

[0027] , , , This is the training parameter matrix for the BiGRU network;

[0028] For the first Bias vectors of each feature map ;

[0029] For BiGRU networks The bias vector at time t;

[0030] and For a moment Information forward and backward propagation GRU unit hidden layer output weights;

[0031] 3) Calculate and update the balance factor that determines the transition between the exploration and development phases in each iteration. and the probability of a whale falling As shown in the following formula:

[0032] ,

[0033] ,

[0034] In the formula, This represents the current iteration number. It changes randomly during each iteration; it determines the current stage of the beluga whale. At that time, DE-BWO was in the exploratory stage; when At that time, DE-BWO was under development;

[0035] 4) Update the beluga whale's position for the next moment based on its position during the exploration phase, as shown in the following formula:

[0036] ,

[0037] In the formula, Indicates the first A beluga whale in Position in dimensions Indicates from The random integer selected in the dimension , , Represents random operators, all of which are Random numbers within a certain range; Indicates the first beluga whale in The updated position on the dimension , The fins of the mirror-image beluga whale face the water's surface. It is an integer;

[0038] 5) When the beluga whale location is in the development stage, the convergence of the BWO algorithm can be improved by introducing a differential evolution strategy;

[0039] (1) Each beluga whale Three random parents were obtained through a differential strategy. The calculation of leader variation is shown in the following formula:

[0040] ,

[0041] In the formula, Indicates the first Individual, , , for Three distinct random numbers are given. , , Indicates a species that is different from an individual in the current population. 3 individuals; for The scaling factor between the two It is the mutated leader individual vector;

[0042] (2) Each beluga leader performs crossover operations on each dimension with a certain probability, resulting in a crossover vector. As shown in the following formula:

[0043] ,

[0044] In the formula, For crossover probability, for A randomly generated integer;

[0045] When the random number is less than the crossover probability, the leader beluga whale undergoes crossover mutation; otherwise, it does not mutate.

[0046] (3) For each vector Calculate the objective function value and adjust its fitness with its parent individuals. The better individuals are selected as the parent individuals for the next generation, as shown in the following formula:

[0047] ,

[0048] Compare the mutated leader position with the fitness value of the current optimal solution. If the mutated leader position is better than the optimal solution, then replace the new solution with the optimal solution and replace the original optimal solution with the new leader position.

[0049] 6) During the whale fall phase, a position update model for the whale fall phase is constructed using the whale's fall step length and the beluga whale's position. The position update model is shown in the following formula:

[0050] ,

[0051] In the formula, , , express Random numbers within a certain range; The step length of the whale's fall can be obtained from the following formula:

[0052] ,

[0053] In the formula, For optimizing the upper limit of weights and parameter variables, For optimizing the lower bounds of weights and parameter variables, The step factor, which relates to population size and the probability of whale decline, can be obtained from the following formula:

[0054] ;

[0055] 7) Determine the relationship between the current number of BWO algorithm iterations and the maximum number of iterations. If the current number of iterations is greater than or equal to the maximum number of iterations, stop the optimization and output the optimal weights and parameters of each network in CNN-BiGRU, and perform model training and validation. Otherwise, return to step 3 to continue the loop iteration.

[0056] 8) Determine the relationship between the current total number of training iterations of the CNN-BiGRU-FCFNN network model and the maximum number of iterations. If the current number of iterations is greater than or equal to the maximum number of iterations, the neural network training stops and the completed CNN-BiGRU-FCFNN network model is output and data verification is performed. Otherwise, return to step two to continue neural network training.

[0057] Furthermore, step five specifically includes the following steps:

[0058] 1) The control data sequence output by the deep learning model is transmitted to the simulated rotor motion equations, which are calculated as follows:

[0059] In grid-connected mode, it is necessary to provide power support to the grid and actively participate in the frequency and voltage regulation of the main grid distribution system. Virtual synchronization control takes output power as the target quantity, which can be expressed as follows:

[0060] ,

[0061] In the formula, , The reference and rated values ​​for the active power output of the photovoltaic energy storage inverter. The active power-frequency droop coefficient of the virtual speed governor. The virtual moment of inertia is denoted as , and is the control variable. This is the virtual damping coefficient. , For virtual rotor angular velocity and rated angular velocity;

[0062] The virtual angular frequency for controlling the photovoltaic energy storage inverter can be obtained from the above formula, as shown in the following equation:

[0063] ,

[0064] When the output voltage of the photovoltaic energy storage inverter deviates normally during system frequency and voltage regulation, the reactive power-voltage droop relationship of the virtual synchronous generator adjusts the reactive power output of the photovoltaic energy storage inverter as shown in the following formula:

[0065] ,

[0066] In the formula, , Reference and rated values ​​for the reactive power output of the photovoltaic energy storage inverter. , These are the actual and rated values ​​of the terminal voltage of the photovoltaic energy storage inverter. The virtual exciter voltage-reactive droop factor;

[0067] 2) Virtual rotor angular velocity output based on rotor equations Power calculation output with grid The vector action time is calculated, and the pulse sequence is obtained to obtain the space vector SVPWM modulation wave signal. The SVPWM signal is used to control the switching of the T-type three-level photovoltaic energy storage inverter to achieve virtual synchronous power output control of the photovoltaic energy storage inverter.

[0068] The present invention has the following beneficial effects:

[0069] 1) The CEEMDAN used in this invention is a powerful signal decomposition tool that can handle complex data signals well. It is superior to traditional methods in handling nonlinear and non-stationary signals, and has the characteristics of good robustness and strong adaptability.

[0070] 2) The CNN-BiGRU-FCFNN network performs efficient feature extraction and classification on photovoltaic energy storage inverter data models, and has strong identification, modeling and prediction capabilities for omnidirectional time operating states, showing excellent performance in nonlinear problems.

[0071] 3) The BWO algorithm based on differential evolution is used to optimize network parameters and improve the prediction accuracy and generalization ability of the CNN-BiGRU-FCFNN model. The BWO algorithm based on differential evolution enhances its local search ability and convergence speed. It has the characteristics of high efficiency, global optimization, avoiding getting trapped in local optima, and simple implementation.

[0072] 4) This invention enables adaptive learning of changes in the virtual synchronous power output of photovoltaic energy storage inverters, which has better robustness and adaptability to complex, high-dimensional grid-connected systems. It can effectively improve the accuracy and control performance of the virtual synchronous power output of photovoltaic energy storage inverters, making them more intelligent and reliable. Attached Figure Description

[0073] Figure 1 A structural diagram of the photovoltaic energy storage power generation system used to implement the method of the present invention.

[0074] Figure 2This is a schematic block diagram of the VSG control system structure of the photovoltaic energy storage inverter used in this invention.

[0075] Figure 3 This is a diagram of the GRU structure used in this invention.

[0076] Figure 4 This is a schematic diagram of the BiGRU structure used in this invention.

[0077] Figure 5 This is a flowchart of the virtual synchronous power output control process for the photovoltaic energy storage inverter used in this invention. Detailed Implementation

[0078] The invention will now be further described with reference to the accompanying drawings.

[0079] As shown in Figure 1, this invention proposes a power control method for a photovoltaic energy storage inverter. The hardware components of the control system that implements this method include: GPU, DSP, ARM processor, VSG control module, and communication module.

[0080] In a pre-trained neural network model control platform, the GPU is primarily used for high-performance inference computation. By predicting new inputs using an existing model and outputting results, it can significantly improve the performance and efficiency of model inference, thereby increasing the responsiveness of the entire application. The DSP focuses on digital signal processing tasks, performing low-power, high-efficiency encoding and decoding. It can also accelerate these operations through highly optimized instruction sets, further enhancing neural network inference performance and reducing power consumption. The ARM processor can run embedded operating systems and software such as Python 3.6 and TensorFlow 1.8.0, providing a reliable execution environment for the model, responding to model inference requests in real time, accelerating model inference, and enabling deep learning technology based on neural networks. The VSG control module uses virtual rotor equations to transform and calculate the control sequence input to the neural network, thereby controlling the output of virtual synchronous power from the photovoltaic energy storage inverter. The communication module transmits data collected from sensors or other devices to the neural network model for processing and transmits the processing results back to the execution device to achieve network control and application. It also employs an optimized EtherCAT network communication protocol to achieve high transmission speed and low communication latency control.

[0081] This invention first uses CEEMDAN to collect current and voltage data at the input and output terminals of a photovoltaic energy storage inverter, and then uses the CEEMDAN algorithm to adaptively process data with different signal-to-noise ratios in order to construct a data model of the photovoltaic energy storage inverter.

[0082] Then, based on the CNN network, feature extraction, data augmentation, and rapid identification of abnormal behavior are performed on the key data of the photovoltaic energy storage inverter data model. Next, based on the BiGRU network, omnidirectional time series modeling is performed on the data features extracted by the CNN network. The output of BiGRU is connected to a Full Connect Neural Network (FCFNN) to output the control sequence.

[0083] The CNN-BiGRU-FCFNN network training simultaneously employs the BWO algorithm, which is an improvement based on Differential Evolution (DE), to optimize the parameters in the network. Finally, the neural network training is completed to obtain a dynamic operation model based on the output of the photovoltaic energy storage inverter, which controls the power output of the photovoltaic energy storage inverter in real time, achieving more accurate and stable power control.

[0084] like Figure 1 , Figure 2 and Figure 5 As shown in the figure, the power control method of a photovoltaic energy storage inverter in this embodiment has the following specific steps:

[0085] The first step is to use the CEEMDAN algorithm to collect current and voltage data at the input and output terminals of the photovoltaic energy storage inverter. The steps are as follows:

[0086] 1) Voltage and current acquisition and processing at the input and output terminals of the photovoltaic energy storage inverter:

[0087] like Figure 2 As shown, at the output of the photovoltaic energy storage inverter, based on the equivalent model of the LCL filter, and neglecting the stray resistance which is very small compared to the inductor impedance, the grid-side inductor current... and the voltage output of the converter The relationship is shown in equation (1):

[0088] (1),

[0089] In the formula, The inductance value is for the LCL filter module that is directly connected to the output terminal of the T-type photovoltaic energy storage inverter. The inductance value of the LCL filter module connected to the power grid. The capacitance value in the LCL filter module; This represents a variable in the complex frequency domain.

[0090] According to equation (1), the voltage and current at the input terminal of the photovoltaic energy storage inverter are collected. , and the three-phase voltage of the capacitor on the LCL filter , , and the three-phase current flowing through the inductor , , and , , and the input current ,Voltage The dq-axis components are obtained by decomposing the photovoltaic energy storage inverter into the synchronous rotating coordinate system. , and , , , .

[0091] The calculation of the output active and reactive power and the effective value of the terminal voltage is shown in the following formula (2):

[0092] (2),

[0093] In the formula, Active power Reactive power This refers to the voltage amplitude at the technical department.

[0094] 2) The CEEMDAN method is used to process the current and voltage data acquired and processed in step 1 above, improving the signal storage accuracy and integrity. The specific steps are as follows:

[0095] (1) Regarding step 1 above, The signal data represents the signal. Repeated run The sub-decomposition operation is assigned to the intrinsic modal components. As shown in equation (3):

[0096] (3),

[0097] (2) Calculate the first residual signal As shown in equation (4):

[0098] (4),

[0099] (3) Operational signals get As shown in equation (5):

[0100] (5),

[0101] In the formula, As the independent variable, The signal is denoised based on the magnitude of the added white noise. The first order of the data signal after decomposition Quantity, It is white noise with a value of 0.

[0102] (4) Calculate the first A residual signal, As shown in equation (6):

[0103] (6),

[0104] (5) Repeat step (3) to calculate the result. indivual The value is as follows (7):

[0105] (7),

[0106] (6) Repeat steps (4) and (5), and the final decomposed original signal is as follows (8):

[0107] (8),

[0108] in, is the independent variable.

[0109] (7) Construct the photovoltaic energy storage inverter data model matrix from the original signals decomposed in step (6) above. .

[0110] 3) The permutation entropy method is used to simplify the data model structure of the photovoltaic energy storage inverter, reduce the computational scale, and improve learning efficiency. The specific steps are as follows:

[0111] (1) For the sequence Reconstruction yields the reconstruction matrix. As shown in equation (9):

[0112] (9),

[0113] In the formula, The dimension of the embedded sequence. The number of reconstructed vectors, For delay time factor, .

[0114] (2) Reconstructing the matrix The row vectors are sorted in ascending order based on the numerical values ​​of their elements to obtain a new matrix. One of the row vectors The number of permutations and combinations of the internal elements is at most 100. kind.

[0115] (3) Calculate in sequence Group row vectors Probability of occurrence And calculate the permutation entropy. As shown in equation (10):

[0116] (10)

[0117] In the formula, For A logarithmic function with base 0.

[0118] (4) The permutation entropy After normalization, the matrix of the new photovoltaic energy storage inverter data model structure is obtained. The normalization is shown in equation (11):

[0119] (11),

[0120] The second step involves using a CNN-BiGRU network to extract data features and train a deep learning model for the simplified photovoltaic energy storage inverter data model. The number of training rounds is from... arrive =10000, the steps are as follows:

[0121] The upper layers of CNN-BiGRU consist of CNNs, which can accept various decomposed variables that affect the virtual synchronous power generation of photovoltaic energy storage inverters. It adopts a 1D CNN consisting of an input layer, a convolutional layer, and a pooling layer.

[0122] 1) Initialize and update the network parameters: the weight matrix of the corresponding convolutional layer , No. Bias vectors of each feature map , .

[0123] 2) Convolutional layers perform convolution operations on multiple input time-series data and then pass the results to the next layer. The specific steps are as follows:

[0124] (1) The convolution operation is used to extract features from time series, as shown in equation (12):

[0125] (12),

[0126] In the formula, For the first The feature output values ​​extracted by each convolutional layer It is the sigmoid activation function; This is a time series matrix of data collected and decomposed for photovoltaic energy storage inverters.

[0127] (2) Use a pooling layer to perform max pooling on the data feature set to obtain new feature data. As shown in equation (13):

[0128] (13)

[0129] In the formula, It is less than the input The pooling size is set to 256. It determines the step size by which the merged region will be moved.

[0130] The lower layers of the CNN-BiGRU network consist of BiGRU networks, which can accept data feature sequences processed by 1D CNN and use them to train the power output model of photovoltaic energy storage inverters, such as... Figure 3 The GRU structure diagram shown uses a BiGRU network, where the GRU network consists of update gates and reset gates. The specific steps are as follows:

[0131] 1) Initialize and update the parameters of the BiGRU network, including the training parameter matrix of the BiGRU network. , , , ,time Information backpropagation GRU unit hidden layer output weights , For a moment Information forward propagation: GRU unit hidden layer output weights, current hidden layer corresponding bias vector .

[0132] 2) The update gate represents the degree of influence of the output information of the hidden layer neuron at the previous time step on the hidden layer neuron at the current time step. The larger the update gate value, the greater the influence. The calculation is shown in the following formula (14):

[0133] (14)

[0134] In the formula, To update the gate output; For the Sigmoid function; For the current input, This represents the output of the hidden layer neurons from the previous time step.

[0135] 3) The reset gate represents the degree to which the output of the hidden layer neuron in the previous time step was ignored. The larger the value of the reset gate, the less information is ignored. The update is shown in the following formula (15):

[0136] (15)

[0137] In the formula, To reset the gate output.

[0138] in It can be calculated from the following formulas (16) and (17):

[0139] (16)

[0140] (17)

[0141] In the formula, This represents the current candidate activation state. It is the hyperbolic tangent function. express and The complex relationship.

[0142] Since the power output control process of photovoltaic energy storage inverters can not only refer to historical data, but is also affected by the parameters of the photovoltaic energy storage inverter at the current moment and the future operating status of the photovoltaic energy storage inverter, the use of BiGRU networks can learn the relationship between the factors affecting power output in the past, present and future and the current power output of the photovoltaic energy storage inverter, and deeply explore the deep feature sequence of the virtual synchronous power output data of the photovoltaic energy storage inverter.

[0143] 4) such as Figure 4 The BiGRU structure shown has its hidden layer at the current time step. From the time propagation forward along the time axis ( Hidden layer output The time propagation along the time axis ( Hidden layer output And the input at the current time The linear superposition of the three factors is shown in equation (18) below:

[0144] (18)

[0145] In the formula, This indicates a gated loop unit.

[0146] The third step involves leveraging the powerful fitting capabilities of the FCFNN network to process the output sequence of the BiGRU neural network to obtain the final output data sequence, as detailed below:

[0147] 1) Initialize and update FCFNN network parameters: weight matrix calculated between neurons Bias vector .

[0148] 2) The FCFNN network consists of an input layer, a hidden layer with a designable number of layers, and an output layer. Each neuron in each layer is connected to all neurons in the next layer to output the virtual synchronous power time series of the photovoltaic energy storage inverter. The relationship between neurons is shown in the following equation (19):

[0149] (19)

[0150] In the formula, The output of the current neuron is derived from the calculations of the neurons in the previous layer. For activation function, This is the input for the neurons in the previous layer that participate in the computation.

[0151] The fourth step is to optimize the parameters of the CNN-BiGRU-FCFNN network using the improved DE-BOW algorithm.

[0152] Because the CNN-BiGRU-FCFNN network has a large number of parameters and a complex nonlinear structure, such as the weights of CNN convolutional layers... , No. Bias vectors of each feature map ( Training parameter matrix of BiGRU network , , , ,time Information forward and backward propagation GRU unit hidden layer output weights , , The bias vector at time t The neurons in the FCFNN network calculate the weight matrix. Bias vector If parameters are simply initialized randomly, redundancy will occur in the algorithm network, resulting in low accuracy. Therefore, an adaptive optimization algorithm is needed to find a suitable solution for parameter optimization.

[0153] Therefore, this invention optimizes the solution capability and recognition accuracy of the CNN-BiGRU-FCFNN network model by using a differential evolution-based improved beluga whale swarm optimization algorithm (DE-BWO). The specific steps are as follows:

[0154] 1) Determine the initial parameters of the DE-BWO optimization algorithm

[0155] Initialize the maximum number of DE-BWO iterations =1000, dimension The total number of parameters and weight matrices to be optimized in the CNN-BiGRU network, and the population size. =100, Leader of the White Whales Beluga whale followers .

[0156] 2) Initialize DE-BWO population location The initial positions of each beluga whale are randomly generated within the parameter setting range, and the population fitness value of the DE-BWO algorithm is determined. As shown in equation (20):

[0157] (20)

[0158] in, The value function of the CNN-BiGRU-FCFNN network parameters. The target value output by the CNN-BiGRU-FCFNN network model;

[0159] These are the weights of the convolutional layers in a CNN network.

[0160] Calculate the weight matrix for the neurons of the FCFNN fully connected neural network;

[0161] , , , This is the training parameter matrix for the BiGRU network;

[0162] For the first Bias vectors of each feature map ;

[0163] For BiGRU networks The bias vector at time t;

[0164] and For a moment Information forward and backward propagation of hidden layer output weights of GRU units.

[0165] 3) Calculate and update the balance factor that determines the transition between the exploration and development phases in each iteration. and the probability of a whale falling As shown in equations (21) and (22):

[0166] (twenty one),

[0167] (twenty two),

[0168] In the formula, This represents the current iteration number. This changes randomly during each iteration. The current stage of the beluga whale is determined, and when... At that time, DE-BWO was in the exploratory stage; when At that time, DE-BWO was under development.

[0169] 4) Update the beluga whale's position for the next moment based on its position during the exploration phase, as shown in equation (23) below:

[0170] (twenty three),

[0171] In the formula, Indicates the first A beluga whale in Position in dimensions Indicates from The random integer selected in the dimension , , Represents random operators, all of which are Random numbers within a certain range; Indicates the first beluga whale in The updated position on the dimension , The fins of the mirror-image beluga whale face the water's surface. It is an integer.

[0172] 5) When the beluga whale location is in the development stage, the convergence of BWO can be improved by introducing a differential evolution strategy.

[0173] (1) Each beluga whale Three random parents were obtained through a differential strategy. The calculation of leader variation is shown in equation (24):

[0174] (twenty four),

[0175] In the formula, Indicates the first Individual, , , for Three distinct random numbers are given. , , Indicates a species that is different from an individual in the current population. The three individuals. for The scaling factor between the two It is the mutated leader individual vector.

[0176] (2) Each beluga leader performs crossover operations on each dimension with a certain probability, resulting in a crossover vector. As shown in equation (25):

[0177] (25)

[0178] In the formula, For crossover probability, for A randomly generated integer.

[0179] When the random number is less than the crossover probability, the leader, the beluga whale, undergoes crossover mutation; otherwise, it does not mutate.

[0180] (3) For each vector Calculate the objective function value and adjust its fitness with its parent individuals. The better individuals are selected as the parent individuals for the next generation, as shown in equation (26) below:

[0181] (26)

[0182] Compare the mutated leader position with the fitness value of the current optimal solution. If the mutated leader position is better than the optimal solution, then replace the optimal solution with the new solution and replace the original optimal solution with the new leader position.

[0183] 6) During the whale fall phase, the position update model for the whale fall phase is constructed using the whale's fall step length and the beluga whale's position. The mathematical model is shown in equation (27) below:

[0184] (27)

[0185] In the formula, , , express A random number within a given range. The step length representing the whale's fall can be obtained from the following formula (28):

[0186] (28)

[0187] In the formula, For optimizing the upper limit of weights and parameter variables, For optimizing the lower bounds of weights and parameter variables, The step factor, which represents the relationship between population size and the probability of whale decline, can be obtained from the following equation (29):

[0188] (29)

[0189] 7) Determine the relationship between the current number of iterations of the White Whale optimization algorithm and the maximum number of iterations. If the current number of iterations is greater than or equal to the maximum number of iterations, the optimization stops and the optimal weights and parameters of each network of CNN-BiGRU are output, and the model is trained and validated. Otherwise, return to step 3 to continue the loop iteration.

[0190] 8) Determine the relationship between the current total number of training iterations of CNN-BiGRU-FCFNN and the maximum number of iterations. If the current number of iterations is greater than or equal to the maximum number of iterations, the neural network training stops and the network model of CNN-BiGRU-FCFNN after training is output and data verification is performed. Otherwise, return to step two to continue neural network training.

[0191] The fifth step involves implanting the successfully trained CNN-BiGRU-FCFNN model into the chip and transmitting the output sequence data to the VSG control module. This module calculates relevant parameter data and ultimately outputs the inverter control signal to control the power output of the photovoltaic energy storage inverter. The specific control steps are as follows:

[0192] 1) The control parameters are transmitted to the simulated rotor motion equations, which are calculated as follows:

[0193] In grid-connected mode, it is necessary to provide power support to the grid and actively participate in system frequency and voltage regulation. Virtual synchronous control takes output power as the target quantity, which can be expressed as shown in the following formula (30):

[0194] (30)

[0195] In the formula, , The reference and rated values ​​for the active power output of the photovoltaic energy storage inverter. The active power-frequency droop coefficient of the virtual speed governor. The virtual moment of inertia is denoted as , and is the control variable. This is the virtual damping coefficient. , For virtual rotor angular velocity and rated angular velocity,

[0196] The virtual angular frequency for controlling the photovoltaic energy storage inverter can be obtained from the above equation (31), as shown in equation (31):

[0197] (31),

[0198] When the output voltage of the photovoltaic energy storage inverter deviates normally during system frequency and voltage regulation, the reactive power-voltage droop relationship of the virtual synchronous generator adjusts the reactive power output of the photovoltaic energy storage inverter as shown in equation (32):

[0199] (32),

[0200] In the formula, , Reference and rated values ​​for the reactive power output of the photovoltaic energy storage inverter. , These are the actual and rated values ​​of the terminal voltage of the photovoltaic energy storage inverter. This represents the virtual exciter voltage-reactive droop coefficient.

[0201] 2) Virtual rotor angular velocity output based on rotor equations Power calculation output with grid The vector action time is calculated, and the pulse sequence is obtained to obtain the space vector SVPWM modulation wave signal. The SVPWM signal is used to control the switching of the T-type three-level photovoltaic energy storage inverter to achieve virtual synchronous power output control of the photovoltaic energy storage inverter.

[0202] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be considered within the scope of protection of the present invention.

Claims

1. A power control method for a photovoltaic energy storage inverter, characterized in that: include Step 1: The CEEMDAN algorithm is used to collect current and voltage data at the input and output terminals of the photovoltaic energy storage inverter, and the CEEMDAN algorithm is used to adaptively process data with different signal-to-noise ratios in order to construct a data model of the photovoltaic energy storage inverter. Step 2: Use a CNN-BiGRU network to extract data features and train deep learning for the photovoltaic energy storage inverter data model; Step 3: Based on the BiGRU neural network, perform omnidirectional time series modeling on the data features extracted by the CNN-BiGRU network, connect the output of the BiGRU neural network with the FCFNN fully connected neural network and input the data sequence to obtain the control data sequence output by the CNN-BiGRU-FCFNN network model; Step 4: Optimize the parameters of the CNN-BiGRU-FCFNN network model using the DE-BWO algorithm based on differential evolution, and finally complete the learning and training of the CNN-BiGRU-FCFNN network model to obtain a deep learning model based on the output control of photovoltaic energy storage inverter. Step 5: Implant the successfully trained deep learning model into the chip module, and output the control data sequence output by the CNN-BiGRU-FCFNN network model in Step 3 through the chip module. After the VSG control module calculates and outputs the control signal, the virtual synchronous power output of the photovoltaic energy storage inverter is controlled. In step two, the upper layer of the CNN-BiGRU network is a CNN network, which is used to accept various decomposed variables that affect the virtual synchronous power generation of the photovoltaic energy storage inverter. The lower layer is a BiGRU network, which is used to receive the data feature sequence after the CNN network is processed, and to train the data model of the photovoltaic energy storage inverter. Step five specifically includes the following steps: 1) The control data sequence output by the deep learning model is transmitted to the simulated rotor motion equations, which are calculated as follows: In grid-connected mode, it is necessary to provide power support to the grid and actively participate in the frequency and voltage regulation of the main grid distribution system. Virtual synchronization control takes output power as the target quantity, which can be expressed as follows: ， In the formula, , The reference and rated values ​​for the active power output of the photovoltaic energy storage inverter. The active power-frequency droop coefficient of the virtual speed governor. The virtual moment of inertia is denoted as , and is the control variable. This is the virtual damping coefficient. , For virtual rotor angular velocity and rated angular velocity; The virtual angular frequency for controlling the photovoltaic energy storage inverter can be obtained from the above formula, as shown in the following equation: ， When the output voltage of the photovoltaic energy storage inverter deviates normally during system frequency and voltage regulation, the reactive power-voltage droop relationship of the virtual synchronous generator adjusts the reactive power output of the photovoltaic energy storage inverter as shown in the following formula: ， In the formula, , Reference and rated values ​​for the reactive power output of the photovoltaic energy storage inverter. , These are the actual and rated values ​​of the terminal voltage of the photovoltaic energy storage inverter. The virtual exciter voltage-reactive droop factor; 2) Virtual rotor angular velocity output based on rotor equations Power calculation output of the grid The vector action time is calculated, and the pulse sequence is obtained to obtain the space vector SVPWM modulation wave signal. The SVPWM signal is used to control the switching of the T-type three-level photovoltaic energy storage inverter to achieve virtual synchronous power output control of the photovoltaic energy storage inverter.

2. The power control method for a photovoltaic energy storage inverter as described in claim 1, characterized in that: Step one specifically includes the following steps: 1) Acquisition and processing of voltage and current data at the input and output terminals of the photovoltaic energy storage inverter; 2) The CEEMDAN algorithm is used to process the voltage and current data in step 1) to improve the signal storage accuracy and integrity; 3) The permutation entropy method is used to simplify the data model of photovoltaic energy storage inverter, reduce the computational scale, and improve learning efficiency.

3. The power control method for a photovoltaic energy storage inverter as described in claim 1, characterized in that: Step four specifically includes the following steps: 1) Determine the initial parameters of the differential evolution-based improved BWO algorithm, i.e., the DE-BWO algorithm; Maximum number of iterations to initialize the DE-BWO algorithm =1000, population size =100, Leader of the White Whales beluga whale followers dimensionality This represents the total number of parameters and weight matrices that need to be optimized in the CNN-BiGRU network. 2) Initialize the population position for the DE-BWO algorithm The initial positions of each beluga whale are randomly generated within the parameter setting range, and the population fitness value of the DE-BWO algorithm is determined. As shown in the following formula: ， in, The value function of the CNN-BiGRU-FCFNN network parameters. The target value output by the CNN-BiGRU-FCFNN network model; These are the weights of the convolutional layers in a CNN network. Calculate the weight matrix for the neurons of the FCFNN fully connected neural network; , , , This is the training parameter matrix for the BiGRU network; For the first Bias vectors of each feature map ; For BiGRU networks The bias vector at time t; and For a moment Information forward and backward propagation GRU unit hidden layer output weights; 3) Calculate and update the balance factor that determines the transition between the exploration and development phases in each iteration. and the probability of a whale falling As shown in the following formula: ， ， In the formula, This represents the current iteration number. It changes randomly during each iteration; it determines the current stage of the beluga whale. At that time, DE-BWO was in the exploratory stage; when At that time, DE-BWO was under development; 4) Update the beluga whale's position for the next moment based on its position during the exploration phase, as shown in the following formula: ， In the formula, Indicates the first A beluga whale in Position in dimensions Indicates from The random integer selected in the dimension , , Represents random operators, all of which are Random numbers within a certain range; Indicates the first beluga whale in The updated position on the dimension , The fins of the mirror-image beluga whale face the water's surface. It is an integer; 5) When the beluga whale location is in the development stage, the convergence of the BWO algorithm can be improved by introducing a differential evolution strategy; (1) Each beluga whale Three random parents were obtained through a differential strategy. The calculation of leader variation is shown in the following formula: ， In the formula, Indicates the first Individual, , , for Three distinct random numbers are given. , , Indicates a species that is different from an individual in the current population. 3 individuals; for The scaling factor between the two It is the mutated leader individual vector; (2) Each beluga leader performs crossover operations on each dimension with a certain probability, resulting in a crossover vector. As shown in the following formula: ， In the formula, For crossover probability, for A randomly generated integer; When the random number is less than the crossover probability, the leader beluga whale undergoes crossover mutation; otherwise, it does not mutate. (3) For each vector Calculate the objective function value and adjust its fitness with its parent individuals. The better individuals are selected as the parent individuals for the next generation, as shown in the following formula: ， Compare the mutated leader position with the fitness value of the current optimal solution. If the mutated leader position is better than the optimal solution, then replace the optimal solution with the new solution and replace the original optimal solution with the new leader position. 6) During the whale fall phase, a position update model for the whale fall phase is constructed using the whale's fall step length and the beluga whale's position. The position update model is shown in the following formula: ， In the formula, , , express A random number within a given range; The step length of the whale's fall can be obtained from the following formula: ， In the formula, For optimizing the upper limit of weights and parameter variables, For optimizing the lower bounds of weights and parameter variables, The step factor, which relates to population size and the probability of whale decline, can be obtained from the following formula: ； 7) Determine the relationship between the current number of BWO algorithm iterations and the maximum number of iterations. If the current number of iterations is greater than or equal to the maximum number of iterations, stop the optimization and output the optimal weights and parameters of each network in CNN-BiGRU, and perform model training and validation. Otherwise, return to step 3 to continue the loop iteration. 8) Determine the relationship between the current total number of training iterations of the CNN-BiGRU-FCFNN network model and the maximum number of iterations. If the current number of iterations is greater than or equal to the maximum number of iterations, the neural network training stops and the completed CNN-BiGRU-FCFNN network model is output and data verification is performed. Otherwise, return to step two to continue neural network training.

Images