CN-121997144-A - Electric power system oscillation mechanism identification method and model construction method

CN121997144ACN 121997144 ACN121997144 ACN 121997144ACN-121997144-A

Abstract

The invention discloses a method for identifying an oscillation mechanism of an electric power system and a method for constructing a model, and relates to the field of low-frequency oscillation of the electric power system; the method comprises the steps of carrying out linear normalization and principal component analysis dimension reduction treatment on data to construct a multi-dimensional time sequence sample to be tested, inputting the sample into a pre-trained deep belief network DBN model, wherein the model is formed by stacking a plurality of limited Boltzmann machines RBM, extracting deep data features through unsupervised layer-by-layer greedy pre-training, and carrying out parameter optimization in combination with supervised fine tuning. The invention can automatically identify negative damping oscillation and forced power oscillation, avoids uncertainty of manually selecting characteristics, effectively inhibits influence of noise and amplitude difference on identification precision, and improves adaptability of the model to actual complex disturbance conditions of the power grid.

Inventors

ZHAO JINGPU
QIN XIAOYUAN
JIA JINGLI
ZHANG CHUNHUI
WANG LIN
JIANG XINXIN
Li Fangjiu

Assignees

三峡金沙江川云水电开发有限公司

Dates

Publication Date: 20260508
Application Date: 20260205

Claims (10)

1. A method for identifying an oscillation mechanism of an electric power system, comprising: step S1, collecting active power data of a generator and active power data of a tie line of a power system; Step S2, preprocessing the acquired data to obtain multidimensional time series data which are used as sample data to be tested, wherein the preprocessing comprises normalization processing and dimension reduction processing; S3, inputting the sample data to be tested into a pre-trained oscillation mechanism identification model, extracting deep features of the sample data to be tested by using the model, classifying the deep features, and outputting an oscillation mechanism identification result, wherein the oscillation mechanism identification result comprises a negative damping oscillation mechanism and a forced power oscillation mechanism; wherein the oscillation mechanism identification model is constructed based on a deep belief network DBN.
2. The method for identifying an oscillation mechanism of an electric power system according to claim 1, wherein the normalizing the data in step S2 by using a linear normalization method comprises: S21, combining the collected active power of the generator and the active power data of the connecting line according to a time sequence to construct multidimensional time sequence data; and S22, processing the multi-dimensional time series data by adopting a linear normalization method, and mapping the data value to the [0,1] interval to obtain the multi-dimensional time series data after the linear normalization processing.
3. The method for identifying an oscillation mechanism of an electric power system according to claim 2, wherein the step S2 of performing the dimension reduction processing on the data by using a principal component analysis method specifically comprises: s23, calculating an average covariance matrix of the multidimensional time series data subjected to linear normalization processing; Step S24, calculating eigenvalues and corresponding normalized eigenvectors of the average covariance matrix; step S25, selecting feature vectors corresponding to the previous p feature values according to a preset variance action rate, and constructing a low-dimensional projection space; and S26, mapping the multi-dimensional time series data subjected to linear normalization processing to the low-dimensional projection space to obtain the multi-dimensional time series data subjected to dimension reduction, and inputting the multi-dimensional time series data serving as the sample data to be tested into the oscillation mechanism identification model.
4. A method of identifying a power system oscillation mechanism according to claim 3, wherein the deep belief network DBN comprises: the input layer is used for receiving the preprocessed multidimensional time series data, and the node number of the input layer is determined by the dimension of the multidimensional time series data; The feature extraction layer is formed by stacking a plurality of limited Boltzmann machines RBMs from bottom to top and is used for self-learning oscillation intrinsic features from the multi-dimensional time sequence data through unsupervised layer-by-layer pre-training, wherein each RBM layer comprises a visual layer and an implicit layer, and a weight matrix is arranged between two adjacent RBMs Connecting; And the top-layer classifier is connected to the top end of the feature extraction layer and is used for classifying the extracted oscillation features based on the fine adjustment of the supervised parameters and outputting corresponding oscillation mechanism class labels.
5. A method for constructing an oscillation mechanism identification model of an electric power system, which is used for constructing the oscillation mechanism identification model of claim 1, characterized by comprising the steps of: The method comprises the steps of M1, constructing a training data set containing an oscillation mechanism label, wherein the training data set contains active power data of a generator and active power data of a connecting wire, and the oscillation mechanism label comprises a negative damping oscillation mechanism and a forced power oscillation mechanism; M2, performing linear normalization processing and dimension reduction processing on the data in the training data set; m3, constructing an initial structure of a Deep Belief Network (DBN), wherein the DBN comprises an input layer, a feature extraction layer and a top classifier, and the feature extraction layer is formed by stacking a plurality of RBMs from bottom to top; M4, based on the multidimensional time series data processed in the step 2, performing layer-by-layer pre-training on the DBN by adopting an unsupervised greedy algorithm, taking the output of the RBM of the current layer as the input of the RBM of the next layer, and training each RBM in turn to initialize the network parameters of the DBN; and M5, performing supervised fine tuning on the pre-trained DBN by using the oscillation mechanism label and adopting a back propagation algorithm, and updating the network parameters to obtain a trained oscillation mechanism identification model.
6. The method for constructing an oscillation mechanism identification model of an electric power system according to claim 5, wherein in the step M4, the unsupervised greedy algorithm trains the boltzmann machine RBM limited in each layer by using a contrast divergence algorithm, and specifically comprises the following steps: Step M41, initializing network parameters of the RBM of the current layer and inputting data Assigning the hidden layer with the visual layer, calculating the activation probability of each hidden unit in the hidden layer by using the Sigmoid activation function, and sampling to obtain the hidden layer state ; Step M42. The hidden layer is laminated As input reverse transfer back to the visual layer, calculating the activation probability of each visual unit in the visual layer by using the network parameters, and sampling to obtain a reconstructed visual layer state ; Step M43, the reconstructed visual layer state is displayed And then the result is used as input to be transferred to the hidden layer again, the activation probability of each hidden unit in the hidden layer is calculated, and the activation probability of the reconstructed hidden layer is obtained ; Step M44, based on the input data Hidden layer state Forward correlation statistics of (a) and the reconstructed visual layer state And reconstructing hidden layer activation probabilities Calculating the gradient and updating the weight matrix, the visual layer bias vector and the hidden layer bias vector of the RBM of the current layer; And step M45, judging whether the reconstruction error meets a preset condition or whether the iteration number reaches a preset maximum value, if so, ending the training of the RBM, otherwise, returning to the step M41 to perform the next iteration.
7. The method for building an oscillation mechanism identification model of a power system according to claim 5, wherein the top classifier is a Softmax classifier, and in step M5, the supervised fine tuning specifically comprises: A forward calculation step, namely inputting the multidimensional time series data processed in the step M2 into the DBN pre-trained in the step M4, extracting the features layer by layer through the feature extraction layer, and taking the hidden layer output of the RBM at the topmost layer as the input of the Softmax classifier; An error calculation step of outputting a predicted probability distribution of each oscillation mechanism category by using the Softmax classifier, and calculating a classification error between the predicted probability distribution and the oscillation mechanism label; and a reverse fine tuning step, namely calculating gradient from top to bottom by utilizing a reverse propagation algorithm based on the classification error, and updating a weight matrix, a visual layer bias vector and an implicit layer bias vector of each layer in the DBN.
8. The method for constructing an oscillation mechanism identification model of a power system according to claim 5, further comprising a parameter setting step for initializing a training hyper-parameter of the DBN before constructing an initial structure of the deep belief network DBN in step M3, the training hyper-parameter comprising: Momentum factor, learning rate, maximum iteration number of single-layer RBM, overall fine-tuning training number of DBN network, and number of hidden layer nodes in RBM of each layer.
9. The method for constructing an oscillation mechanism identification model of an electric power system according to claim 5, wherein in step M1, the determination of the oscillation mechanism label is based on the following: The negative damping oscillation mechanism is that the upper envelope curve characteristic of the power oscillation curve is in a concave shape, and the label of the corresponding oscillation mechanism is marked as 0; The upper envelope curve characteristic of the forced power oscillation mechanism is in an upward convex shape or a straight line shape, and the label of the corresponding oscillation mechanism is marked as 1.
10. An electronic device, comprising: and a memory communicatively coupled to the at least one processor; wherein the memory stores instructions executable by the at least one processor, by executing the instructions stored by the memory, causing the at least one processor to perform the method of any one of claims 1-4.

Description

Electric power system oscillation mechanism identification method and model construction method Technical Field The invention relates to the field of low-frequency oscillation of a power system, in particular to a power system oscillation mechanism identification method and a model construction method. Background The statements in this section merely provide background information related to the present disclosure and may not constitute prior art. The low frequency oscillation is mainly studied in two major categories, namely negative damped free oscillation and forced oscillation. The treatment measures for these two oscillations are different. When negative damping free oscillation occurs, the oscillation will subside by adjusting the running mode and increasing the system damping. When forced oscillation occurs, the oscillation can subside only by clearing the disturbance source. Therefore, an accurate oscillation type discrimination result is an important basis for the low-frequency oscillation emergency control assistance decision. At present, researches on low-frequency oscillation type identification have formed various technical routes, and can be mainly divided into the following categories: 1. Methods based on model resolution and feature comparison such methods start from the physical nature of the oscillation. The unique features of the forced oscillation, which are composed of the superposition of the free component and the forced component, are revealed, for example, by deriving an analytical expression of the forced oscillation, and are distinguished by analyzing the response components (such as damping ratio). Based on WAMS measured data, the system compares dynamic characteristic differences of two types of oscillations in the whole process. 2. The method based on the signal time domain waveform characteristics emphasizes the morphological characteristics of the oscillation waveform, including extremum change law, envelope curve shape and the like in the oscillation starting stage, so as to realize rapid discrimination. For example, a secondary difference method is utilized, rapid discrimination is realized in more than ten cycles after oscillation is started by analyzing the sign of a power extreme point differential sequence, the form of an oscillation envelope is focused, the envelope is extracted by utilizing Hilbert-Huang transform, the concave-convex characteristic recognition is carried out by adopting a support vector machine, and the comprehensive judgment is carried out by analyzing the amplitude change rule and the damping ratio obtained by EMD decomposition. 3. The method is based on frequency domain analysis, and is characterized in that the frequency spectrum characteristics of signals are researched, and the difference of different oscillations in the frequency domain is utilized for identification, so that the method generally has strong anti-noise interference capability. The above methods all take a certain different characteristic of free oscillation and forced oscillation as the basis of type discrimination, although the theory is strict, the requirement on oscillation measurement data is considered to be strict, the actual oscillation data is sometimes difficult to obtain, and the ideal effect is difficult to obtain in the practical system application. In addition, if the system damping is very small, the system damping is in a weak damping or near undamped state, at the moment, all state quantity changes are expressed as constant-amplitude oscillation, and characteristic information such as a single-slave oscillation phenomenon cannot be distinguished. And when the system is frequently disturbed randomly, the problems of noise, multi-frequency aliasing and the like are frequently accompanied in the actual measurement signal, and the direct application of the method can cause the reduction of the identification accuracy of the oscillation mechanism. Disclosure of Invention The invention aims to solve the problems in the prior art, provides a power system oscillation mechanism identification method and a model construction method, and solves the following technical problems: 1. The method solves the problems that the model-dependent technology has poor adaptability caused by topology/parameter change in a novel power system and oscillation types cannot be distinguished under a weak damping working condition; 2. The problem that the positioning accuracy of an oscillation source is low in non-ideal data environments such as noise, multi-frequency aliasing and data missing in a measurement data driving technology is solved; 3. Constructing an end-to-end mechanism identification framework, uniformly modeling and training a negative damping oscillation sample and a forced power oscillation sample, and realizing mechanism type discrimination based on supervised fine tuning; 4. The normalization and characteristic reconstruction mechanism is adopted, so that the influence of no