A Subsynchronous Oscillation Mode Identification Method Based on SVM and LSTM

Abstract

This invention discloses a subsynchronous oscillation mode identification method based on SVM and LSTM, comprising the following steps: S1: acquiring voltage and current signals of the power system; S2: preprocessing the voltage and current signals to obtain subsynchronous oscillation signals, dividing them into training sample data and test sample data; S3: constructing and training a long short-term memory neural network model; S4: constructing and training a support vector machine model; S5: fusing the trained long short-term memory neural network model and the support vector machine model to obtain a combined model; S6: inputting the subsynchronous oscillation signal to be tested into the combined model to complete the subsynchronous oscillation mode identification. This subsynchronous oscillation mode identification method combines support vector machines and long short-term memory neural networks to identify relevant parameters of subsynchronous oscillations, improving the accuracy of identification and providing assurance for power system stability.

Description

Technical Field

[0001] This invention belongs to the field of subsynchronous oscillation mode identification technology in power systems, specifically relating to a subsynchronous oscillation mode identification method based on SVM and LSTM. Background Technology

[0002] With the large-scale integration of new energy sources such as wind power into the grid and the widespread use of power electronic equipment, modern power systems are gradually evolving towards high-efficiency and high-power systems. This has led to increasingly prominent subsynchronous oscillation problems, exhibiting characteristics such as strong time-varying nature, strong nonlinearity, multimodal operation, and wide-area propagation. Subsynchronous oscillations pose significant challenges to the safe and stable operation of power system equipment; therefore, accurately identifying subsynchronous oscillation parameters and suppressing them is crucial.

[0003] Currently, most subsynchronous oscillation mode identification methods are based on signal processing techniques, such as Fast Fourier Transform (FFT), Prony Transform, Synchronous Compressed Wavelet Transform (SST), and Hilbert-Huang Transform (HHT). FFT and Prony Transform can only analyze linear stationary signals, wavelet transform is not ideal for signals with time-varying characteristics, and Hilbert-Huang can analyze nonlinear, non-stationary signals, but its empirical mode decomposition suffers from mode aliasing. With the continuous development of artificial intelligence, its low dependence on models, powerful learning ability for complex nonlinear relationships between large amounts of data, and rapid adaptability to stochastic and time-varying environments have enhanced its capabilities. Core AI algorithms such as reinforcement learning and machine learning are also being applied to the analysis of broadband oscillation problems, demonstrating certain advantages. Summary of the Invention

[0004] To address the above problems, this invention proposes a subsynchronous oscillation mode identification method based on SVM and LSTM.

[0005] The technical solution of this invention is: a subsynchronous oscillation mode identification method based on SVM and LSTM, comprising the following steps:

[0006] S1: Acquires voltage and current signals from the power system;

[0007] S2: Preprocess the voltage and current signals to obtain the subsynchronous oscillation signal, and divide the subsynchronous oscillation signal into training sample data and test sample data;

[0008] S3: Construct and train a long short-term memory neural network model using training sample data;

[0009] S4: Build and train a support vector machine model using test sample data;

[0010] S5: The trained Long Short-Term Memory Neural Network model and the trained Support Vector Machine model are fused to obtain a combined model;

[0011] S6: Input the subsynchronous oscillation signal to be tested into the combined model to complete the subsynchronous oscillation mode identification.

[0012] Further, in step S2, the voltage signal and the current signal are filtered to obtain the subsynchronous oscillation signal. 80% of the subsynchronous oscillation signal is used as training sample data, and 20% of the subsynchronous oscillation signal is used as test sample data.

[0013] Furthermore, step S3 includes the following sub-steps:

[0014] S31: Tag for generating subsynchronous oscillation signals;

[0015] S32: Construct a long short-term memory neural network model based on the labels of the subsynchronous oscillation signals;

[0016] S33: Input the training sample data into the long short-term memory neural network model to train the long short-term memory neural network model.

[0017] Furthermore, in step S31, the label of the subsynchronous oscillation signal includes a frequency label and an attenuation factor label;

[0018] The frequency tag generation method is as follows: divide the frequency of the subsynchronous oscillation signal into equal parts and add a frequency tag to the frequency of each segment of the subsynchronous oscillation signal;

[0019] The method for generating the attenuation factor label is as follows: divide the subsynchronous oscillation signal into weakly damped and negatively damped signals, and add an attenuation factor label to the subsynchronous oscillation signal.

[0020] Furthermore, in step S32, the long short-term memory neural network model includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, and a fully connected layer connected in sequence.

[0021] The loss function expression for a fully connected layer is:

[0022]

[0023] In the formula, w represents the overshoot parameter, α represents the frequency value of the subsynchronous oscillation signal, β represents the damping value of the subsynchronous oscillation signal, L represents the number of neurons in the first convolutional layer, G represents the number of neurons in the first pooling layer, P represents the number of neurons in the second convolutional layer, Q represents the number of neurons in the second pooling layer, and σ l σ represents the weight of the l-th neuron in the first convolutional layer. g σ represents the weight of the g-th neuron in the first pooling layer. pσ represents the weight of the p-th neuron in the second convolutional layer. q This represents the weight of the q-th neuron in the second pooling layer.

[0024] Furthermore, in step S33, the specific method for training the long short-term memory neural network model is as follows: the gradient descent method is used to adjust the weight parameters of the fully connected layer in the direction of the negative gradient until the loss function value of the fully connected layer is less than the set loss threshold.

[0025] Furthermore, in step S4, the expression for the objective function F of training the support vector machine model is:

[0026]

[0027] In the formula, w represents the weight coefficient of the learning optimal hyperplane, b represents the bias coefficient of the learning optimal hyperplane, n represents the number of slack variables, C represents the penalty factor, and ξ represents the weight coefficient of the learning optimal hyperplane. i Let i represent the i-th slack variable. Let y represent the vector of the i-th slack variable, and y represent the test sample data.

[0028] Furthermore, in step S5, the expression for the combined model minE is:

[0029]

[0030]

[0031] In the formula, n represents the number of slack variables, x i w represents the true value of the i-th slack variable. k Indicates the weights of the combined model. represents the predicted value of the combined model, K represents the number of model combinations, and Loss represents the loss function of the fully connected layer.

[0032] Further, in step S6, the subsynchronous oscillation signal to be tested is input into the combined model to generate a subsynchronous oscillation model. The amplitude, phase angle, frequency and attenuation factor of the subsynchronous oscillation signal to be tested are obtained according to the subsynchronous oscillation model. The waveform is fitted according to the amplitude, phase angle, frequency and attenuation factor of the subsynchronous oscillation signal to be tested to complete the subsynchronous oscillation mode identification.

[0033] Furthermore, the expression for the subsynchronous oscillation model x(t) is:

[0034]

[0035] In the formula, A m σ represents the amplitude of the m-th synchronous oscillation signal to be measured. m f represents the attenuation factor of the m-th synchronous oscillation signal to be measured. mThis represents the frequency of the m-th synchronous oscillation signal to be measured. Let t represent the phase angle of the m-th synchronous oscillation signal to be measured, and t represent time.

[0036] The beneficial effects of this invention are: the subsynchronous oscillation mode identification method combines support vector machine and long short-term memory neural network, giving full play to the advantages of each algorithm, and identifies the relevant parameters of subsynchronous oscillation through the combined model of the two, thereby improving the accuracy of identification, providing stronger support for subsequent analysis of subsynchronous oscillation, and ensuring the stability of power system. Attached Figure Description

[0037] Figure 1 This is a flowchart of a subsynchronous oscillation mode identification method based on SVM and LSTM. Detailed Implementation

[0038] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0039] like Figure 1 As shown, this invention provides a subsynchronous oscillation mode identification method based on SVM and LSTM, comprising the following steps:

[0040] S1: Acquires voltage and current signals from the power system;

[0041] S2: Preprocess the voltage and current signals to obtain the subsynchronous oscillation signal, and divide the subsynchronous oscillation signal into training sample data and test sample data;

[0042] S3: Construct and train a long short-term memory neural network model using training sample data;

[0043] S4: Build and train a support vector machine model using test sample data;

[0044] S5: The trained Long Short-Term Memory Neural Network model and the trained Support Vector Machine model are fused to obtain a combined model;

[0045] S6: Input the subsynchronous oscillation signal to be tested into the combined model to complete the subsynchronous oscillation mode identification.

[0046] In step S1, the voltage signal and current signal are the energy signals of the subsynchronous oscillation point to be estimated. The collected energy signal is a time series composed of instantaneous scalar values ​​within a specific time length, such as a sampling time of 10s and a sampling frequency of 100Hz.

[0047] In this embodiment of the invention, in step S2, the voltage signal and the current signal are filtered to obtain the subsynchronous oscillation signal. 80% of the subsynchronous oscillation signal is used as training sample data, and 20% of the subsynchronous oscillation signal is used as test sample data.

[0048] In this embodiment of the invention, step S3 includes the following sub-steps:

[0049] S31: Tag for generating subsynchronous oscillation signals;

[0050] S32: Construct a long short-term memory neural network model based on the labels of the subsynchronous oscillation signals;

[0051] S33: Input the training sample data into the long short-term memory neural network model to train the long short-term memory neural network model.

[0052] In this embodiment of the invention, the input and output of the previous layer are used as the input and output of the long short-term memory neural network model for training according to the weight parameters between the layers of the long short-term memory neural network. The model parameters are adjusted and the cost function is the squared reconstruction error. After training is completed, the identification results are output by the long short-term memory neural network model using voltage and current signal data. The obtained identification results are compared with the true values, and the relative error is used as the measurement standard to measure the identification accuracy of the LSTM network model.

[0053] In this embodiment of the invention, in step S31, the label of the subsynchronous oscillation signal includes a frequency label and an attenuation factor label.

[0054] The frequency tag generation method is as follows: divide the frequency of the subsynchronous oscillation signal into equal parts and add a frequency tag to the frequency of each segment of the subsynchronous oscillation signal;

[0055] The method for generating the attenuation factor label is as follows: divide the subsynchronous oscillation signal into weakly damped and negatively damped signals, and add an attenuation factor label to the subsynchronous oscillation signal.

[0056] In this embodiment of the invention, in step S32, the long short-term memory neural network model includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, and a fully connected layer connected in sequence.

[0057] The loss function expression for a fully connected layer is:

[0058]

[0059] In the formula, w represents the overshoot parameter, α represents the frequency value of the subsynchronous oscillation signal, β represents the damping value of the subsynchronous oscillation signal, L represents the number of neurons in the first convolutional layer, G represents the number of neurons in the first pooling layer, P represents the number of neurons in the second convolutional layer, Q represents the number of neurons in the second pooling layer, and σ l σ represents the weight of the l-th neuron in the first convolutional layer. g σ represents the weight of the g-th neuron in the first pooling layer. p σ represents the weight of the p-th neuron in the second convolutional layer. q This represents the weight of the q-th neuron in the second pooling layer.

[0060] In this embodiment of the invention, the specific method for training the long short-term memory neural network model in step S33 is as follows: the gradient descent method is used to adjust the weight parameters of the fully connected layer in the direction of the negative gradient until the loss function value of the fully connected layer is less than the set loss threshold.

[0061] In this embodiment of the invention, in step S4, the expression for the objective function F of training the support vector machine model is:

[0062]

[0063] In the formula, w represents the weight coefficient of the learning optimal hyperplane, b represents the bias coefficient of the learning optimal hyperplane, n represents the number of slack variables, C represents the penalty factor, and ξ represents the weight coefficient of the learning optimal hyperplane. i Let i represent the i-th slack variable. Let y represent the vector of the i-th slack variable, and y represent the test sample data.

[0064] Support vector machine (SVM) regression was used to find the best-fit curve and obtain the relevant parameters of each mode. Subsynchronous mode identification involves given a set of time-varying measurements and fitting these measurements to a pre-defined set of time-varying waveforms, minimizing the error between the actual measured waveform and the obtained waveform. Therefore, SVM was used to fit the amplitude, frequency, phase angle, and attenuation factor of the subsynchronous oscillation mode. The SVM model assumes that a maximum error of ε between the predicted and true values ​​can be tolerated; the loss is calculated only when the absolute value of the difference between the predicted and true values ​​is greater than 2ε. ε was set to 0.01.

[0065] In this embodiment of the invention, in step S5, the expression for the combined model minE is:

[0066]

[0067]

[0068] In the formula, n represents the number of slack variables, x iw represents the true value of the i-th slack variable. k Indicates the weights of the combined model.

[0069] represents the predicted value of the combined model, K represents the number of model combinations, and Loss represents the loss function of the fully connected layer.

[0070] In this embodiment of the invention, in step S6, the subsynchronous oscillation signal to be tested is input into the combined model to generate a subsynchronous oscillation model. The amplitude, phase angle, frequency and attenuation factor of the subsynchronous oscillation signal to be tested are obtained according to the subsynchronous oscillation model. The waveform is fitted according to the amplitude, phase angle, frequency and attenuation factor of the subsynchronous oscillation signal to be tested to complete the subsynchronous oscillation mode identification.

[0071] In this embodiment of the invention, the expression for the subsynchronous oscillation model x(t) is:

[0072]

[0073] In the formula, A m σ represents the amplitude of the m-th synchronous oscillation signal to be measured. m f represents the attenuation factor of the m-th synchronous oscillation signal to be measured. m This represents the frequency of the m-th synchronous oscillation signal to be measured. Let t represent the phase angle of the m-th synchronous oscillation signal to be measured, and t represent time.

[0074] In this embodiment of the invention, taking a synthesized signal as an example, the subsynchronous oscillation model is set to first order and the relevant parameters are A1 = 0.2f1 = 25, σ1 = 0.05. The identified data is input into the trained model. The Long Short-Term Memory Neural Network model estimates relevant parameters by constructing a classifier, the Support Vector Machine model obtains relevant parameters by fitting the subsynchronous oscillation waveform using regression, and the combined model obtains the final identification result by weight allocation using least squares method. The obtained identification result parameters are A1 = 0.2, f1 = 25.04, and σ1 = 0.05. The frequency error was 0.16%, and there were no errors in other parameters.

[0075] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for sub-synchronous oscillation mode identification based on SVM and LSTM, characterized in that, The method comprises the following steps: S1: collecting voltage signals and current signals of a power system; S2: preprocessing the voltage signals and the current signals to obtain a subsynchronous oscillation signal, and dividing the subsynchronous oscillation signal into training sample data and test sample data; S3: constructing and training a long short-term memory neural network model by using the training sample data; S4: constructing and training a support vector machine model by using the test sample data; S5: fusing the trained long short-term memory neural network model and the trained support vector machine model to obtain a combined model; S6: inputting a to-be-detected subsynchronous oscillation signal into the combined model to complete subsynchronous oscillation mode identification. The step S3 comprises the following sub-steps: S31: generating a label of the subsynchronous oscillation signal; S32: constructing a long short-term memory neural network model according to the label of the subsynchronous oscillation signal; S33: inputting the training sample data into the long short-term memory neural network model to train the long short-term memory neural network model. In the step S32, the long short-term memory neural network model comprises a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer and a fully connected layer connected in sequence. The loss function Loss of the fully connected layer is expressed as: In the formula, w represents a tuning parameter, a represents a frequency value of the subsynchronous oscillation signal, β represents a damping value of the subsynchronous oscillation signal, L represents a number of neurons of the first convolutional layer, G represents a number of neurons of the first pooling layer, P represents a number of neurons of the second convolutional layer, Q represents a number of neurons of the second pooling layer, σ l represents a weight of the lth neuron in the first convolutional layer, σ g represents a weight of the gth neuron in the first pooling layer, σ p represents a weight of the pth neuron in the second convolutional layer, σ q represents a weight of the qth neuron in the second pooling layer.

2. The method of claim 1, wherein the method is characterized by, In the step S2, the voltage signals and the current signals are filtered to obtain the subsynchronous oscillation signal, 80% of the subsynchronous oscillation signal is taken as the training sample data, and 20% of the subsynchronous oscillation signal is taken as the test sample data.

3. The SVM and LSTM based subsynchronous oscillation mode identification method of claim 1, wherein, In the step S31, the label of the subsynchronous oscillation signal comprises a frequency label and an attenuation factor label. The frequency label is generated by equally dividing the frequency of the subsynchronous oscillation signal and adding a frequency label to the frequency of each segment of the subsynchronous oscillation signal. The attenuation factor label is generated by dividing the subsynchronous oscillation signal into weak damping and negative damping and adding an attenuation factor label to the subsynchronous oscillation signal.

4. The SVM and LSTM based subsynchronous oscillation mode identification method of claim 1, wherein, In the step S33, the specific method for training the long short-term memory neural network model is to adjust the weight parameters of the fully connected layer in the negative gradient direction by using the gradient descent method until the loss function value of the fully connected layer is less than a set loss threshold.

5. The SVM and LSTM based subsynchronous oscillation mode identification method of claim 1, wherein, In the step S4, the expression of the objective function F of the support vector machine model is: In the formula, w represents a weight coefficient of learning an optimal hyperplane, b represents a bias coefficient of learning an optimal hyperplane, n represents a number of slack variables, C represents a penalty factor, and ξ i represents the i-th slack variable, represents a vector of the i-th slack variable, and y represents test sample data.

6. The SVM and LSTM based subsynchronous oscillation mode identification method of claim 1, wherein, In the step S5, the expression of the combined model minE is: where n represents the number of relaxation variables, x i represents the true value of the i-th relaxation variable, w k represents the combination model weight, represents the combination model predicted value, K represents the number of model combinations, and Loss represents the loss function of the fully connected layer.

7. The SVM and LSTM based subsynchronous oscillation mode identification method of claim 1, wherein, In the step S6, the to-be-detected subsynchronous oscillation signal is input into the combined model to generate a subsynchronous oscillation model, the amplitude, the phase angle, the frequency and the attenuation factor of the to-be-detected subsynchronous oscillation signal are obtained according to the subsynchronous oscillation model, and the waveform is fitted according to the amplitude, the phase angle, the frequency and the attenuation factor of the to-be-detected subsynchronous oscillation signal to complete the subsynchronous oscillation mode identification.

8. The method of claim 7, wherein the method is characterized by, The expression of the subsynchronous oscillation model x(t) is: In the formula, A m denotes the amplitude of the mth to-be-detected subsynchronous oscillation signal, σ m denotes the damping factor of the mth to-be-detected subsynchronous oscillation signal, f m denotes the frequency of the mth to-be-detected subsynchronous oscillation signal, denotes the phase angle of the mth to-be-detected subsynchronous oscillation signal, and t denotes time.

Images