CN-122022010-A - Neural network frequency response model and method based on physical information

CN122022010ACN 122022010 ACN122022010 ACN 122022010ACN-122022010-A

Abstract

The invention discloses a frequency response model and a frequency response method based on a physical information neural network, which belong to the technical field of power system frequency prediction. And normalizing various coefficients of the generator frequency curves, the inertial center frequency of the power system, the unbalanced active power and the frequency response model under different scenes and different disturbance to obtain training data. The output of the physical information neural network includes the coefficients of the frequency response model transfer function and the unbalanced active power. The output of the physical information neural network is substituted into a frequency response model containing unknown parameters, so that the frequency change of the power system under various disturbance scenes can be rapidly and accurately predicted, and the frequency safety and stability of the complex power system are enhanced.

Inventors

SUN ZHENGLONG
JIA QINGLIN
ZHANG RUI
PAN CHAO
LI ZHENXIN
YU YUANZHI
CAI GUOWEI

Assignees

东北电力大学
国网吉林省电力有限公司吉林供电公司

Dates

Publication Date: 20260512
Application Date: 20251230

Claims (7)

1. A frequency response model based on a physical information neural network, comprising: Reading frequency information, unbalanced active power and frequency response model parameters of the power system as training data and normalizing the training data to ensure consistency and comparability of the data; The method comprises the steps of designing a physical information neural network structure, namely accurately configuring the quantity and connection modes of neurons of an input layer, a middle layer and an output layer, selecting proper optimizers such as SGD, adam and RMSprop to improve the efficiency and accuracy of network training, setting a learning rate to balance the convergence speed in the training process and the generalization capability of a model, determining iteration times to ensure that the network reaches a preset performance standard in a sufficient training period, combining a physical item loss function with a data item loss function, wherein the physical item loss function is the mean square error of a frequency response model and the inertia center frequency of an actual power system, and the data item loss function is the mean square error of the output value of the neural network and the frequency response module coefficient; And inputting the characteristic quantity output by the physical information neural network into a frequency response model, and generating and outputting a complete power system frequency curve through calculation of the model.
2. The frequency response model based on the physical information neural network according to claim 1, wherein the training data is parameters of each generator frequency, unbalanced active power and frequency response model for the first 2 seconds of the power system, and the parameters are normalized for training and prediction processes, and the frequency sampling step length of each generator is 0.01.
3. The frequency response model based on the physical information neural network according to claim 2 is characterized in that the neural network adopts a fully-connected neural network architecture, an input layer receives 10 characteristic quantities, comprises 5 hidden layers, is provided with 64 neurons in each layer, an output layer outputs 5 characteristic quantities, two transposed layers are embedded and an Adam optimizer is configured, the network optimizer adopts the Adam optimizer, the learning rate is 0.00004, the iteration number is 20, and the structural design aims at optimizing the data processing capacity of the network and the accuracy of output results.
4. The method for generating a frequency response model based on a physical information neural network according to claim 3, wherein the physical term loss function is a transfer function of the frequency response model: the output parameters A, B, C, D and delta P of the physical information neural network are substituted into a physical item loss function, a frequency response curve is generated by the model, the frequency response curve is compared with the inertial center frequency of the power system, a physical item loss value is obtained by calculating the mean square error between the frequency response curve and the inertial center frequency, the prediction accuracy of the model is evaluated by the data item loss function through comparing the mean square error of the output parameters of the neural network and the frequency response model parameters, and the total loss function of the model is formed by the sum of the physical item loss function and the data item loss function.
5. The frequency response model based on the physical information neural network according to claim 4, wherein the feature quantity output by the neural network is converted back to an original scale from a normalized state, the output parameters are inversely normalized, the normalized feature quantity output by the neural network is mapped back to an actual value thereof by reversing a normalization transformation previously applied to the training data, and the inversely normalizing step ensures that an output result of the neural network is directly compared with the unprocessed actual data and is used for subsequent physical information analysis.
6. The frequency response model based on the physical information neural network according to claim 5, wherein the physical information neural network outputs the frequency response model parameters and the unbalanced active power predicted values and substitutes the frequency response model parameters and the unbalanced active power predicted values into a transfer function of the frequency response model, and the frequency response curve can be generated by the model through the process, so that the prediction and analysis of the dynamic frequency of the power system are realized.
7. A method of using the physical information neural network-based frequency response model of any one of claims 1 to 6, comprising: Step one, acquiring frequency curves of all generators under different scenes and different disturbance, calculating the inertia center frequency of a power system and active power disturbance of the power system as training data; secondly, reading a csv file containing a response observation value, carrying out normalization processing on the frequency information of the power system and the frequency response model data, and scaling the frequency response data to a uniform range by normalization, wherein the processed data is used for training and predicting processes; step three, inputting the training data obtained in the step two into a physical information neural network for training; and fourthly, inputting the test data into a trained model, outputting the test data as frequency response model parameters and unbalanced active power predicted values, substituting the predicted parameters into the frequency response model, and comparing the frequency prediction with the inertial center frequency of the actual power system.

Description

Neural network frequency response model and method based on physical information Technical Field The invention relates to the field of power system frequency response models, in particular to a frequency response model and method based on a physical information neural network. Background At present, along with the advancement of a double-carbon strategy, the development of new energy power generation becomes an essential factor for accelerating the reduction of carbon emission and guiding the innovation of green technology, and the continuous advancement of industrial structures and energy structure adjustment. The energy structure is improved, meanwhile, the problems of weak disturbance resistance, reduced inertia, high uncertainty of new energy power generation and the like are also brought to the system, and when the system is greatly disturbed, the system can be subjected to rapid frequency drop and even frequency collapse. Therefore, the method can rapidly and accurately predict the frequency of the power system and has important significance for the safety protection of the power system. When the working condition of the power system is complex and changeable, the traditional frequency response model is difficult to rapidly and accurately predict the frequency curve of the power system, so that the frequency stability control cannot be timely performed. Therefore, the power system frequency curve is rapidly and accurately predicted, and the method has important significance for improving the frequency stability of the power system and preventing the system frequency from collapsing. Disclosure of Invention Aiming at the problems, the invention provides the frequency response model and the method based on the physical information neural network, which can rapidly and accurately predict the inertial center frequency curve of the power system when disturbance occurs in different scenes, and improve the frequency safety and stability of the complex power system. The frequency response model based on the physical information neural network comprises the steps of reading power system frequency, unbalanced active power and frequency response model parameters and normalizing the parameters to ensure consistency and comparability of data, designing the structure of the physical information neural network, including precisely configuring the quantity and connection modes of neurons of an input layer, a middle layer and an output layer, selecting proper optimizers such as SGD, adam or RMSprop to improve the efficiency and accuracy of network training, setting a proper learning rate to balance convergence speed in the training process and generalization capability of the model, and determining iteration times to ensure that the network reaches a preset performance standard in a sufficient training period. The loss function of the physical information neural network is combined with the data item loss function by adopting a physical item loss function; the physical item loss function is the mean square error of the frequency response model and the inertial center frequency of the power system, and the data item loss function is the mean square error of the neural network output value and the frequency response model coefficient; performing inverse normalization on the characteristic quantity output by the neural network; The characteristic quantity output by the physical information neural network is input into a frequency response model, a complete power system frequency curve is generated and output through calculation of the model, and the input data quantity is the parameters of each generator frequency, unbalanced active power of the system and the frequency response model 2 seconds before the power system, and normalization processing is carried out for training and prediction processes. The frequency sampling step length of the generator is 0.01, so that the processed data can better adapt to the requirements of the model, and the efficiency and performance of the model are improved. The neural network adopts a fully-connected neural network architecture. The input layer of the network is designed to receive 10 characteristic quantities, i.e. frequency data representing 10 generators in the first two seconds. The network comprises 5 hidden layers, each hidden layer has 64 neurons, the output layer comprises 5 characteristic quantities, in addition, two transposed layers are embedded in the network, the function of the two transposed layers is to ensure that the data tensor is correct in the transmission process among all layers of the network and the accuracy of a final output format is ensured, an optimizer adopts an Adam optimizer, the learning rate is 0.00004, and the iteration number is 20. Such a structural design aims at optimizing the data processing capacity of the network and the accuracy of the output result; The physical term loss function is the transfer function of the frequency response model: