CN-121981162-A - Model training method and energy storage system virtual impedance self-adaptive adjusting method

CN121981162ACN 121981162 ACN121981162 ACN 121981162ACN-121981162-A

Abstract

The application discloses a model training method and an energy storage system virtual impedance self-adaptive adjustment method, wherein the model training method uses a circulating neural network as a model framework, takes a state vector of a system as input, the training network can output fault types and virtual impedance adjustment amounts according to an input state vector sequence, the trained circulating neural network is used as a virtual impedance adjustment model, real-time state vectors of the system are acquired in actual application, the trained virtual impedance adjustment model is input to generate a predicted virtual impedance adjustment amount, and virtual impedance at the next moment is calculated according to the virtual impedance at the current moment and the virtual impedance adjustment amount, so that self-adaptive adjustment of the virtual impedance is realized, and stable operation of the system is ensured.

Inventors

TANG QI
HE SHENGHONG
ZHONG WEI
ZHANG ZHEMING
ZHENG WEIQIN
YE XIAOGANG

Assignees

广东电网有限责任公司佛山供电局

Dates

Publication Date: 20260505
Application Date: 20260127

Claims (10)

1. The self-adaptive adjustment method for the virtual impedance of the energy storage system is characterized by comprising the following steps of: modeling a state space of the system to generate a state space model of the system; acquiring a state vector of the current moment of the system according to the state space model, wherein the state vector comprises voltage, current and virtual impedance; Inputting the state vector at the current moment into a virtual impedance adjustment model, and predicting and generating a fault type and a virtual impedance adjustment quantity; And calculating the virtual impedance at the next moment according to the virtual impedance at the current moment and the virtual impedance adjustment quantity.
2. The method of claim 1, wherein calculating the virtual impedance at the next time based on the virtual impedance at the current time and the virtual impedance adjustment amount comprises: the following adaptive adjustment relation of virtual impedance is constructed: Wherein, the Representing the virtual impedance at time t +1, The virtual impedance at time t is indicated, Represents the virtual impedance adjustment amount predicted by the virtual impedance adjustment model, Represent learning rate Determined from the voltage sag amplitude and harmonic distortion rate, Representing the gradient of the virtual impedance adjustment model, 。
3. The method of claim 2, wherein the learning rate The following calculation is determined according to the voltage drop amplitude and the harmonic distortion rate: , Wherein, the The basic learning rate is represented by the number of learning, Representing the magnitude of the voltage sag, The harmonic distortion rate is represented by the ratio of the harmonic distortion, Representing the maximum allowable harmonic distortion rate, Representing a nominal voltage reference.
4. A model training method for training a virtual impedance adjustment model according to any one of claims 1 to 3, for predicting a virtual impedance adjustment amount, comprising: Acquiring a historical state vector sequence of a system, wherein the historical state vector sequence is generated by modeling a state space of the system, the historical state vector sequence comprises state vectors which are continuous in time, and the state vectors comprise voltage, current and virtual impedance; The historical state vector sequence is used as input of a cyclic neural network, and the cyclic neural network is trained to output fault types and virtual impedance adjustment amounts according to the input state vector sequence; And constructing a loss function according to the output of the cyclic neural network, updating network parameters of the cyclic neural network by using the loss function, and taking the trained cyclic neural network as a virtual impedance adjustment model.
5. The method of claim 4, wherein the recurrent neural network includes an input layer, a first recurrent layer, a second recurrent layer, a fully connected layer, and an output layer according to a direction of feature delivery.
6. The method of claim 4, wherein constructing a loss function from the output of the recurrent neural network comprises: Wherein, the Representing a loss of classification, Representing the virtual impedance target value, Representing the predicted value of the virtual impedance adjustment amount output by the recurrent neural network, , Represents the positive sequence virtual impedance adjustment quantity, Represents the negative sequence virtual impedance adjustment quantity, Represents the zero sequence virtual impedance adjustment quantity, And And (5) representing a weight coefficient, wherein the weight coefficient is adjusted according to the loss value in the training process of the cyclic neural network.
7. The method of claim 6, wherein the adjustment of the weight coefficients is represented by: Wherein, the And The weight coefficient representing the s-th training round, Representing the loss of classification for the s-1 training round, The regression loss for the s-1 training round is shown.
8. A model training apparatus for training a virtual impedance adjustment model, comprising: The system comprises a state vector acquisition module, a state vector generation module and a state vector generation module, wherein the state vector acquisition module is used for acquiring a historical state vector sequence of the system, the historical state vector sequence is generated by carrying out state space modeling on the system, and the state vector comprises voltage, current and virtual impedance; The network training module is used for taking the historical state vector sequence as the input of the cyclic neural network, training the cyclic neural network, and outputting fault types and virtual impedance adjustment amounts according to the input state vector sequence; And the network updating module is used for constructing a loss function according to the output of the cyclic neural network, updating the network parameters of the cyclic neural network by using the loss function, and taking the trained cyclic neural network as a virtual impedance adjustment model.
9. An energy storage system virtual impedance self-adaptive adjustment device, comprising: The state space modeling module is used for carrying out state space modeling on the system and generating a state space model of the system; The state vector acquisition module is used for acquiring a state vector of the current moment of the system according to the state space model, wherein the state vector comprises voltage, current and virtual impedance; a virtual impedance prediction module for inputting the state vector at the current time into a virtual impedance adjustment model trained by the model training device of claim 8, and predicting the generated fault type and the virtual impedance adjustment amount; And the virtual impedance adjusting module is used for calculating the virtual impedance at the next moment according to the virtual impedance at the current moment and the virtual impedance adjusting quantity.
10. An electronic device comprising a memory storing computer executable instructions and a processor, which when executed by the processor causes the device to perform the energy storage system virtual impedance adaptive adjustment method of any one of claims 1-3 and/or the model training method of any one of claims 4-7.

Description

Model training method and energy storage system virtual impedance self-adaptive adjusting method Technical Field The invention belongs to the control technology of an energy storage system, and particularly relates to a model training method and an energy storage system virtual impedance self-adaptive adjustment method. Background With the continuous development of modern power systems to distributed energy sources and smart grids, the role of energy storage systems in the grids is increasingly highlighted. In order to improve the stability and the electric energy quality of the system, a virtual impedance technology is widely adopted in recent years, and the voltage, the current characteristics and the power distribution of a power grid are adjusted by introducing virtual impedance in a control link. The traditional virtual impedance control method generally depends on a state space model established based on physical parameters and power grid topology, and the model can describe the voltage and current dynamic relationship among nodes in the system more accurately. However, most of the existing methods predefine virtual impedance parameters in the design stage, and adopt a fixed or semi-fixed control strategy, which limits the response speed and adaptability of the system to complex working conditions such as load fluctuation, fault disturbance, instantaneous power grid frequency deviation and the like to a certain extent. Disclosure of Invention Based on the above, the invention aims to provide a model training method and an energy storage system virtual impedance self-adaptive adjustment method, which describe the dynamic behavior of the system by using a state space model, and utilize a cyclic neural network to carry out self-adaptive adjustment of virtual impedance, so that the system can monitor the state of the system in real time, thereby dynamically adjusting impedance parameters. In a first aspect, the present invention provides a model training method, the trained model being used for predicting a virtual impedance adjustment amount, comprising: Acquiring a historical state vector sequence of the system, wherein the historical state vector sequence is generated by modeling a state space of the system, and the state vector comprises voltage, current and virtual impedance; The historical state vector sequence is used as input of the cyclic neural network, and the cyclic neural network is trained to output fault types and virtual impedance adjustment amounts according to the input state vector sequence; And constructing a loss function according to the output of the cyclic neural network, updating network parameters of the cyclic neural network by using the loss function, and taking the trained cyclic neural network as a virtual impedance adjustment model. Further, the recurrent neural network comprises an input layer, a first recurrent layer, a second recurrent layer, a fully connected layer and an output layer according to the characteristic transfer direction. Further, constructing a loss function from the output of the recurrent neural network includes: Wherein, the Representing a loss of classification,Representing the virtual impedance target value,Representing the predicted value of the virtual impedance adjustment amount output by the recurrent neural network,,Represents the positive sequence virtual impedance adjustment quantity,Represents the negative sequence virtual impedance adjustment quantity,Represents the zero sequence virtual impedance adjustment quantity,AndAnd (5) representing a weight coefficient, wherein the weight coefficient is adjusted according to the loss value in the training process of the cyclic neural network. Further, the adjustment of the weight coefficient is expressed as follows: Wherein, the AndThe weight coefficient representing the s-th training round,Representing the loss of classification for the s-1 training round,The regression loss for the s-1 training round is shown. Further, the recurrent neural network adopts an LSTM network. In a second aspect, the present invention provides a method for adaptively adjusting virtual impedance of an energy storage system, including: modeling a state space of the system to generate a state space model of the system; Acquiring a state vector of the system at the current moment according to a state space model, wherein the state vector comprises voltage, current and virtual impedance; inputting a state vector at the current moment into a virtual impedance adjustment model trained by the method of the first aspect, and predicting and generating a fault type and a virtual impedance adjustment quantity; And calculating the virtual impedance at the next moment according to the virtual impedance at the current moment and the virtual impedance adjustment quantity. Further, calculating the virtual impedance at the next time according to the virtual impedance at the current time and the virtual impedance adjustment amount includ