Power system low frequency oscillation category recognition method based on multi-physical characteristic fusion and deep learning

Abstract

The application discloses a power system low-frequency oscillation category identification method based on multi-physical feature fusion and deep learning, collects an oscillation signal of a power system, constructs a time-frequency energy spectrum and a phase plane trajectory diagram of the oscillation signal, constructs a hybrid neural network model, inputs double-channel features into the hybrid neural network model for training, and obtains a trained hybrid neural network model; after feature extraction, the oscillation signal to be identified is input into the trained oscillation category identification model, and an identification result of forced oscillation or free oscillation is obtained. The application effectively solves the problem that single time-domain or frequency-domain features in a traditional method are difficult to distinguish forced oscillation from free oscillation, improves the distinguishability of features, overcomes the problem that a traditional signal processing method is easily affected by mode aliasing and feature ambiguity when processing non-stationary and strong noise interference signals, and overcomes the limitation that a single data-driven model lacks deep fusion of system dynamics mechanism.

Description

Technical Field

[0001] This invention relates to the field of power system oscillation category identification methods, specifically a power system low-frequency oscillation category identification method based on multi-physical feature fusion and deep learning. Background Technology

[0002] With the large-scale integration of new power electronic equipment such as grid-connected wind turbines, the dynamic characteristics of power systems are becoming increasingly complex, and low-frequency oscillations in power systems have become a major risk threatening the safe and stable operation of the power grid.

[0003] Low-frequency oscillations in power systems can be categorized into forced oscillations and free oscillations.

[0004] Forced oscillations are typically triggered by external periodic disturbances, requiring the location and removal of these disturbances for effective suppression. Free oscillations, on the other hand, arise from insufficient system damping, necessitating intervention through adjusting system parameters or implementing damping control. Since their mechanisms differ and the suppression measures required vary, accurate and rapid identification of the oscillation type is a crucial prerequisite for developing targeted control strategies.

[0005] However, in actual power grid operation, forced oscillations and free oscillations exhibit similar waveforms in the time domain, making them difficult to distinguish effectively based solely on the oscillation phenomena themselves. Furthermore, measured signals are often accompanied by random noise and nonlinear interference, leading to confusion and misjudgment using traditional identification methods based on single time-domain features or spectral analysis. Therefore, constructing a high-precision and robust low-frequency oscillation category identification method is of significant practical and theoretical value for ensuring the safe and stable operation of the power system. Summary of the Invention This invention provides a method for identifying low-frequency oscillation categories in power systems based on multi-physical feature fusion and deep learning, in order to solve the problem that existing technologies make it difficult to identify low-frequency oscillation categories in power systems.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for identifying low-frequency oscillation categories in power systems based on multi-physical feature fusion and deep learning is as follows:

[0008] Step 1: Obtain low-frequency oscillation signal sequences of the power system over multiple time periods. The true oscillation category of the low-frequency oscillation signal sequence for each time period is known, and the true oscillation category is one of forced oscillation or free oscillation.

[0009] Construct time-domain energy maps and phase plane trajectory maps of low-frequency oscillation signal sequences for each time period, use the time-domain energy maps and phase plane trajectory maps of low-frequency oscillation signal sequences for each time period as dual-channel feature samples, and form a training set from the dual-channel feature samples of each time period;

[0010] Step 2: Generate a hybrid neural network model and train the hybrid neural network model using the dual-channel fused feature map samples in the training set obtained in Step 1.

[0011] During training, a hybrid neural network model is used to obtain the corresponding dual-channel fusion feature map based on each dual-channel fusion feature map sample. Then, a three-dimensional fusion feature tensor is extracted from the dual-channel fusion feature map. Subsequently, the dual-channel fusion feature map is reshaped into a feature sequence, and the dynamic evolution law of the feature sequence is explored. Finally, linear transformation and activation are performed based on the dynamic evolution results to obtain the final predicted oscillation category of each dual-channel fusion feature map sample.

[0012] During training, the parameters of the hybrid neural network model are updated based on the calculation results of the loss function, and the trained hybrid neural network model is used as a low-frequency oscillation category recognition model.

[0013] Step 3: Obtain the time-domain energy spectrum and phase plane trajectory map of the low-frequency oscillation signal sequence to be identified in the power system, and form the current dual-channel feature sample; input the current dual-channel feature sample into the low-frequency oscillation category identification model, and the low-frequency oscillation category identification model outputs the oscillation category identification result corresponding to the current dual-channel feature sample.

[0014] Furthermore, in step 1, wavelet transform is performed on the low-frequency oscillation signal sequence for each time period to obtain the time-domain energy map corresponding to the low-frequency oscillation signal sequence for each time period.

[0015] Furthermore, in step 1, phase plane analysis is performed on the low-frequency oscillation signal sequence for each time period to obtain the phase plane trajectory diagram of the low-frequency oscillation signal sequence for each time period.

[0016] Furthermore, in step 2, the hybrid neural network model includes a dual-channel input layer, a CNN spatial feature extractor, a feature reshaping layer, an LSTM temporal analyzer, and a classification layer;

[0017] During training, a dual-channel fusion feature map is obtained based on each dual-channel fusion feature map sample through a dual-channel input layer; a three-dimensional fusion feature tensor is extracted from the dual-channel fusion feature map through a CNN spatial feature extractor; the dual-channel fusion feature map is reshaped into a feature sequence through a feature reshaping layer; the dynamic evolution pattern of the feature sequence is mined through an LSTM time series analyzer; finally, a classification layer performs linear transformation and activation based on the dynamic evolution results to obtain the final predicted oscillation category of each dual-channel fusion feature map sample.

[0018] Furthermore, the CNN spatial feature extractor includes three cascaded convolutional blocks, each of which includes a learnable convolutional kernel, batch normalization, ReLU activation function, and max pooling. The first convolutional block takes each dual-channel fused feature map as input, the second convolutional block takes the output feature map obtained from the first convolutional block as input, and the third convolutional block takes the output feature map obtained from the second convolutional block as input.

[0019] In each convolutional block, a sliding window operation is performed on the input feature map using a learnable filter to extract local spatial features. Then, batch normalization, ReLU activation, and max pooling are applied to the local spatial features to obtain the output feature map. Finally, the three-dimensional fused feature tensor is obtained by the third convolutional block.

[0020] Furthermore, the feature reshaping layer performs serialized slicing and flattening of the three-dimensional fused feature tensor along the time axis to obtain a feature sequence.

[0021] Furthermore, the LSTM time series analyzer includes multiple cascaded LSTM layers. The feature sequence is input into the first LSTM layer, and the dynamic evolution pattern is mined step by step through multiple LSTM layers to obtain the hidden state vector of the last time step as the dynamic evolution result of the feature sequence.

[0022] Furthermore, the classification layer includes a fully connected layer and a Softmax activation function. A two-dimensional vector is obtained by linear transformation of the dynamic evolution result of the classification layer. Then, the two-dimensional vector is activated by the Softmax activation function to obtain the two-dimensional probability distribution of the predicted oscillation category for each dual-channel feature sample. The predicted oscillation category with the larger probability value is taken as the final predicted oscillation category for each dual-channel feature sample.

[0023] Furthermore, the update process for the parameters of the hybrid neural network model adopts a staged training strategy.

[0024] In the first stage, dual-channel feature samples are input into the CNN spatial feature extractor, and all trainable parameters of the CNN spatial feature extractor are updated using stochastic gradient descent.

[0025] In the second stage, the 3D fusion feature tensor output by the CNN spatial feature extractor is reshaped and then input into the LSTM temporal analyzer. The Adam optimizer is used to update all trainable parameters of the LSTM temporal analyzer and the fully connected layers in the classification layer.

[0026] Furthermore, in step 2, the loss function during training is calculated as the average cross-entropy loss of each dual-channel feature sample during training of the hybrid neural network model.

[0027] Compared with the prior art, the advantages of the present invention are:

[0028] This invention addresses the problem of low-frequency oscillation classification in power systems, proposing a method based on multi-physical feature fusion and deep learning. First, a dual-channel feature extraction framework is constructed, fusing continuous wavelet transform time-frequency energy spectrum and phase plane trajectory diagram, transforming the time-frequency domain features and dynamic stability features of the oscillation signal into visualized two-dimensional image features. Based on this, a "spatiotemporal joint" deep learning architecture is designed, utilizing convolutional neural networks to extract the spatial texture and geometric topological features of the oscillation signal, and employing long short-term memory networks to deeply mine the dynamic evolution logic of features over time, achieving high-precision and robust identification of forced oscillations and free oscillations.

[0029] This invention utilizes a dual-channel feature extraction framework that integrates the time-frequency energy spectrum of continuous wavelet transform with the phase plane trajectory diagram. This framework transforms the time-frequency domain features and dynamic stability features of oscillating signals into visualized two-dimensional image features, effectively solving the problem that traditional methods struggle to distinguish between forced oscillations and free oscillations using only time-domain or frequency-domain features, thus improving the discriminability of the features.

[0030] This invention utilizes a spatiotemporal joint deep learning architecture composed of a convolutional neural network and a long short-term memory network to extract spatial texture and geometric topological features from dual-channel features, and deeply explores the dynamic evolution logic of features over time. This overcomes the problems of traditional signal processing methods being susceptible to modal mixing and feature ambiguity when dealing with non-stationary signals and strong noise interference, as well as the limitations of single data-driven models that lack deep integration of system dynamics mechanisms, and significantly improves the accuracy and robustness of the identification model.

[0031] This invention has significant advantages in both identification accuracy and anti-interference capability, and is scientifically sound and highly applicable. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the method in an embodiment of the present invention.

[0033] Figure 2 This is a diagram of the CNN spatial feature extraction network architecture in the hybrid neural network model of this invention.

[0034] Figure 3 This is a schematic diagram illustrating the principle of feature reshaping and LSTM temporal feature extraction in the hybrid neural network model of this invention.

[0035] Figure 4 Topology diagram of an improved IEEE 4-unit two-zone system model for wind turbines with a grid structure.

[0036] Figure 5 This is a comparison chart showing the accuracy of the method in this embodiment with other methods under different noise conditions. Detailed Implementation

[0037] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0038] like Figure 1 As shown in the figure, this embodiment discloses a method for identifying low-frequency oscillation categories in power systems based on multi-physical feature fusion and deep learning. The process is as follows:

[0039] Step 1: Obtain low-frequency oscillation signal sequences from multiple time periods of the power system. Each time period's low-frequency oscillation signal sequence contains the low-frequency oscillation signals at each discrete time sampling point within that time period. The true oscillation category of each time period's low-frequency oscillation signal sequence is known, and it can be either forced oscillation or free oscillation. Preprocess the low-frequency oscillation signal sequences for each time period, including missing value imputation and normalization, to obtain the preprocessed low-frequency oscillation signals for each time period. Then, construct the time-domain energy map and phase plane trajectory map of the preprocessed low-frequency oscillation signal sequences for each time period. Use the time-domain energy map and phase plane trajectory map of the low-frequency oscillation signal sequences for each time period as dual-channel feature samples, and form a training set from the dual-channel feature samples of each time period.

[0040] In this embodiment, the time-domain energy spectrum corresponding to the low-frequency oscillation signal sequence of each time period is obtained by performing wavelet transform on the low-frequency oscillation signal sequence of each time period. The specific process is as follows:

[0041] (A1) Determine the mother wavelet function

[0042] In this embodiment, the complex Morlet wavelet function is selected as the mother wavelet function for the transform. As shown in the following formula:

[0043]

[0044] In the formula: t is the time variable, f b f is the bandwidth parameter of the wavelet. c The center frequency parameter of the wavelet; e is the imaginary unit, and e is the natural base.

[0045] (A2) Taking any time period as the target time period, let the low-frequency oscillation signal sequence of the target time period be... ,in The sampling interval for discrete-time sampling points. This is the sampling point number.

[0046] A set of time-shift factors and a set of scale factors are defined. Since the low-frequency oscillation signal sequence is a discrete time sequence, this embodiment uses all discrete time sampling points in the low-frequency oscillation signal sequence of the target time period as the set of time-shift factors for the target time period. Furthermore, this embodiment sets the set of scale factors using a binary discretization method.

[0047] For time shift factors in the time shift factor set and the scale factor in the scale factor set By combining pairs of these pairs, multiple time-shift scale combinations are obtained. , ).

[0048] Then, based on each combination of time shift scales ( , ), respectively for the mother wavelet function By performing scaling and translation, each combination of time shift scales is obtained. , The corresponding wavelet basis function As shown in the following formula:

[0049]

[0050] In the formula: It is a scale factor; This is the time shift factor. , ∈R, and ≠0.

[0051] (A3) Sample each discrete time point in the low-frequency oscillation signal sequence of the target time period. Substitute each time-shift scale combination into ( , wavelet basis functions (i.e., substitution) The parameter t in the text is used to obtain each combination of time shift scales ( , The sampled values ​​of the wavelet basis function at each discrete time sampling point in the target time period. and by all sampled values Each combination of time shift scales constitutes ( , The wavelet basis function sampled value sequence.

[0052] (A4) Combine each time shift scale ( , The wavelet basis function sampled value sequence and the low-frequency oscillation signal sequence of the target time period are respectively subjected to inner product operation to obtain the corresponding time shift scale combination. , The corresponding wavelet coefficients are shown in the following formula:

[0053]

[0054] In the formula: For each time shift scale combination ( , The corresponding wavelet coefficients; For each time shift scale combination ( , The sampled values ​​of the complex conjugate of the wavelet basis function.

[0055] Then, for each combination of time shift scales ( , The wavelet coefficients corresponding to each time shift scale are squared, thus obtaining the combination of each time shift scale ( ). , The corresponding oscillation signal energy value.

[0056] (A5) Construct a two-dimensional coordinate grid with the discrete time sampling points corresponding to the time shift factor as the horizontal axis and the scale factor as the vertical axis; map the energy value of the oscillation signal corresponding to each coordinate point in the grid to different color depths or gray levels, thereby obtaining the time domain energy spectrum corresponding to the low frequency oscillation signal of the target time period.

[0057] In this embodiment, phase plane analysis is performed on the low-frequency oscillation signal sequence for each time period to obtain the phase plane trajectory diagram of the low-frequency oscillation signal sequence for each time period. The specific process is as follows:

[0058] Suppose that in the low-frequency oscillation signal sequence of the target time period, the oscillation signal value at any discrete time sampling point t is... .

[0059] First, the low-frequency oscillation dynamic process of the power system is described using a second-order differential equation. In the second-order differential equation for the low-frequency oscillation dynamic process of the power system, the oscillation signal value is used as the basis for the equation. Let the state variable be as shown in the following equation:

[0060]

[0061] In the formula: For the system damping ratio, The frequency of undamped free oscillation. The external disturbance source is the trajectory on the phase plane, which is the state point. The path of evolution over time t, where State variables The first derivative with respect to time.

[0062] If the oscillation type corresponding to the low-frequency oscillation signal sequence in the target time period is forced oscillation, forced oscillation usually occurs in areas with positive damping ( In systems with these characteristics, there are continuous periodic external disturbance sources. Stimulation, among which The amplitude of the external periodic disturbance source represents the magnitude of the disturbance intensity. ω represents the angular frequency of the external periodic disturbance source, indicating the rate of change of the disturbance.

[0063] According to linear system theory, after the transient component decays, the steady-state response of the system exhibits a constant-amplitude oscillation with the same frequency as the disturbance source, as shown in the following equation:

[0064]

[0065] In the formula: Let be the steady-state amplitude of the forced oscillation. The phase angle of the oscillating signal.

[0066] After eliminating the time variable t, the phase trajectory equation corresponding to the low-frequency oscillation signal sequence of the target time period when the oscillation category is forced oscillation can be obtained, as shown in the following equation:

[0067]

[0068] The above equation shows that the trajectory of forced oscillations on the phase plane is a closed ellipse or circle. From a physical perspective, this corresponds to a stability limit cycle in nonlinear dynamics. At this point, the energy injected by the external disturbance and the energy dissipated by the system's positive damping reach a dynamic equilibrium within one cycle, and the system state is constrained to a closed trajectory.

[0069] Finally, based on the phase trajectory equation corresponding to the low-frequency oscillation signal sequence of the target time period when the oscillation category is forced oscillation, a phase trajectory plane diagram corresponding to the low-frequency oscillation signal sequence of the target time period when the oscillation category is forced oscillation is drawn.

[0070] If the low-frequency oscillation signal sequence in the target time period corresponds to an oscillation category of negative damped oscillation in free oscillation, the negative damped oscillation originates from a change in system parameters that causes the damping ratio to become negative. At this point, the homogeneous solution of the second-order differential equation of the low-frequency oscillation dynamic process of the power system exhibits a divergent form, as shown in the following equation:

[0071]

[0072] In the formula: The initial amplitude of the oscillation. The angular frequency of the damped oscillation. The initial phase angle, It is a natural constant.

[0073] Based on the solution of the homogeneous equation, the phase trajectory plane diagram corresponding to the low-frequency oscillation signal sequence of the target time period when the oscillation category is negative damped oscillation in free oscillation can be drawn.

[0074] because The oscillation amplitude envelope grows exponentially with time. On the phase plane, the state point... The distance to the origin increases monotonically with time:

[0075]

[0076] Therefore, when the oscillation category is negatively damped oscillation within free oscillation, the phase trajectory corresponding to the low-frequency oscillation signal sequence in the target time period exhibits a logarithmic spiral that continuously spreads outward from the equilibrium point. In the dynamic topological classification, this equilibrium point belongs to an unstable focus. Its physical essence is that the system continuously generates self-excited energy, leading to the continuous accumulation and divergence of the system's total energy.

[0077] If the low-frequency oscillation signal sequence corresponding to the target time period corresponds to a weakly damped oscillation in free oscillation, the weakly damped oscillation occurs when the system damping ratio is positive but extremely small ( This type of oscillation is typically triggered by a step disturbance or fault impact on the system and is considered a free response of the system. The dynamic equation for low-frequency oscillation signal sequences of the weakly damped oscillation category is similar to that of negatively damped oscillations, also a homogeneous differential equation, as shown below:

[0078]

[0079] because The time-domain general solution of the homogeneous differential equation for the weakly damped oscillation category is in the form of a damped oscillation, as shown in the following equation:

[0080]

[0081] Based on the time-domain general solution of the homogeneous differential equation of the dynamics for weakly damped oscillations, the phase trajectory plane diagram corresponding to the low-frequency oscillation signal sequence of the target time period for weakly damped oscillations in the category of free oscillations can be plotted.

[0082] At this point, the exponent term It decreases monotonically with time. Therefore, in the phase plane, when the oscillation type is a weakly damped oscillation in a free oscillation, the distance from the state point in the phase trajectory corresponding to the low-frequency oscillation signal sequence of the target time period to the origin gradually decreases with time, as shown in the following equation:

[0083]

[0084] If the oscillation category corresponding to the low-frequency oscillation signal sequence in the target time period is a strongly damped oscillation in free oscillation ( When the initial state is reached, the system is in an overdamped state, with two unequal negative real roots as its characteristic roots. The time-domain response is the sum of two exponentially decaying terms, and the phase trajectory shows that it directly approaches the equilibrium point from the initial state point along a non-oscillating path.

[0085] If the oscillation category corresponding to the low-frequency oscillation signal sequence in the target time period is critical damped oscillation in free oscillation ( When the phase trajectory is at the critical damping state, the characteristic roots are equal negative real roots, and the phase trajectory also converges directly to the origin.

[0086] Furthermore, a few undamped oscillations (ζ=0) exist within free oscillations, originating from the lack of internal damping and energy conservation within the system; forced oscillations arise from the dynamic equilibrium reached between continuous external energy injection and damping dissipation. Although their physical mechanisms differ, their phase diagram geometry is similar, which is precisely the difficulty and value of distinguishing between them using deep learning in this embodiment.

[0087] In this embodiment, the time-domain energy map and phase plane trajectory map of the low-frequency oscillation signal sequence for each time period are used to form the dual-channel feature samples for that time period, and the dual-channel feature samples of each time period are used to form a training set.

[0088] Step 2: Generate a hybrid neural network model. The hybrid neural network model is trained in batches using the dual-channel fused feature map samples from the training set obtained in Step 1, resulting in a trained hybrid neural network model. Each batch uses several dual-channel fused feature map samples to train the hybrid neural network model.

[0089] The hybrid neural network model in this embodiment consists of a dual-channel input layer, a CNN spatial feature extractor, a feature reshaping layer, an LSTM temporal analyzer, and a classification layer.

[0090] In this embodiment, during each batch of training, each dual-channel feature sample is first input into the dual-channel input layer. The dual-channel input layer uses the time-domain energy map as the first channel and the phase plane trajectory map as the second channel. The two channels are independent of each other. The time-domain energy map and the phase plane trajectory map are spliced ​​together in the channel dimension to obtain the dual-channel fused feature map corresponding to the dual-channel feature sample. The dual-channel fused feature map includes the time-domain energy map of the first channel and the phase plane trajectory map of the second channel.

[0091] In this embodiment, a three-dimensional fusion feature tensor X is extracted from each dual-channel fusion feature map using a CNN spatial feature extractor. fea .like Figure 2 As shown, the CNN spatial feature extractor consists of three cascaded convolutional blocks. The first convolutional block includes blocks of size 5. A learnable convolutional kernel (i.e., filter) with 16 channels, batch normalization (BN), non-linear activation function ReLU, and a 2 The first convolutional block uses MaxPool with a stride of 2. The second convolutional block contains blocks of size 3. A learnable convolutional kernel (i.e., filter) with 32 channels, batch normalization (BN), non-linear activation function ReLU, and a 2 The first convolutional block uses MaxPool with a stride of 2. The second convolutional block contains blocks of size 3. A learnable convolutional kernel (i.e., filter) with 64 channels, batch normalization (BN), non-linear activation function ReLU, and a 4-channel learnable convolutional kernel (i.e., filter). Step size is 4 MaxPool.

[0092] The first convolutional block takes each dual-channel fused feature map as input, the second convolutional block takes the output feature map obtained from the first convolutional block as input, and the third convolutional block takes the output feature map obtained from the second convolutional block as input.

[0093] In each convolutional block, a sliding window operation is performed on the input feature map using a learnable filter to extract local spatial features. These local spatial features are then batch normalized, ReLU activated, and max-pooled to obtain the output feature map. For the k-th output feature map obtained from the l-th convolutional block, the convolution operation can be expressed as:

[0094]

[0095] In the formula: Let be the activation value of the k-th output feature map of the l-th convolutional block at position (i, j); This represents the pixel value of the c-th channel in the output feature map of the (l-1)-th convolutional block (i.e., the previous level), at the spatial location (i+u, j+v), which is the value on the input feature map of the current convolutional block; This represents the weight parameters of the k-th convolutional kernel in the l-th convolutional block at offset (u, v), corresponding to the c-th input channel. This indicates the number of channels in the input feature map of the current convolutional block; It is represented as the size of the convolution kernel in the height direction; This indicates the size of the convolution kernel in the width direction; This represents the offset index of the convolution kernel in the height direction; This represents the offset index of the convolution kernel in the width direction; Indicates the index of the input channel; It is represented as the bias term corresponding to the k-th convolution kernel in the l-th convolution block.

[0096] The third convolutional block of the CNN spatial feature extractor outputs a 3D fused feature tensor X. fea .

[0097] In this embodiment, the 3D fused feature tensor X obtained by the CNN spatial feature extractor is processed by the feature reshaping layer. feat Serialization slicing and flattening are performed along the time axis to obtain a feature sequence S={x1,x2,…,x...} of length 56. 56 This is the output of the feature reshaping layer. The specific processing procedure of the feature reshaping layer is as follows:

[0098] Keeping the time axis dimension unchanged at 56, the three-dimensional fused feature tensor X is fused along the time axis. feat The time slice is divided into 56 time slices. At each time step t, data of all frequency axis dimensions (14) and feature channel numbers (64) at that time are extracted and flattened into a one-dimensional vector x. t Then each x t The dimension is 14 × 64 = 896, and the calculation formula for this process is:

[0099]

[0100] In the formula: This represents taking all frequency axes and feature channel data of the feature tensor at time step t; Flatten represents the flattening operation; D = 14 × 64 = 896 is the feature vector dimension for each time step; R is the real number field.

[0101] Therefore, the feature reshaping layer yields the feature sequence S={x1, x2, ..., x...} 56}, which serves as the input to the LSTM timing analyzer.

[0102] In this embodiment, the dynamic evolution pattern of the feature sequence S is mined by an LSTM time series analyzer, and the LSTM time series analyzer outputs the hidden state vector h of the last time step to the classification layer. t This is the result of the dynamic evolution of the feature sequence S.

[0103] like Figure 3As shown, the LSTM time series analyzer consists of multiple cascaded LSTM layers. The first LSTM layer takes the feature sequence S as input, and each LSTM layer except the first takes the output of the previous LSTM layer as input. Each LSTM layer dynamically adjusts the transmission and updating of sequence information by introducing cell states as a carrier of long-term information. The output gate of the LSTM layer is used to calculate the output state at the current time step to determine the amount of memory information to be released at the current time step. The specific calculation formula is as follows:

[0104]

[0105]

[0106] In the formula, It is the weight matrix of the output gate; It is the bias vector; It is the memory unit at the current time step t; This indicates that the hidden state of the previous time step is concatenated with the input of the current time step; ( ) represents the Sigmoid activation function; This represents the output vector of the output gate at time step t; This represents the hidden state vector at time step t, and also serves as the output vector of the LSTM layer at that time step.

[0107] In this embodiment, the classification layer includes a fully connected layer and a Softmax activation function. The classification layer is used to process the hidden state vector h of the last time step output by the LSTM. t A linear transformation is performed to obtain a two-dimensional vector. Then, the two-dimensional vector is activated by the Softmax activation function, thereby obtaining the two-dimensional probability distribution of the predicted oscillation category for each dual-channel feature sample. The predicted oscillation category with the larger probability value is taken as the final predicted oscillation category for each dual-channel feature sample. The predicted oscillation category is either forced oscillation or free oscillation.

[0108] The formula for calculating the linear transformation of the fully connected layer is as follows:

[0109] z=W fc ·h t +b fc

[0110] In the formula: W fc b is the weight matrix of the fully connected layer; fc Let z be the bias vector; z is a two-dimensional vector, z = [z1, z2]. T ,

[0111] z1 and z2 are the logic values ​​for the forced oscillation and free oscillation categories, respectively.

[0112] The formula for calculating the Softmax activation function is:

[0113]

[0114] In the formula: It is a two-dimensional probability distribution. , These are the predicted probabilities of belonging to forced oscillation and free oscillation, respectively. ; As a category, =1 indicates forced oscillation. =2 indicates free oscillation.

[0115] During training, the parameters of the hybrid neural network model are updated based on the loss function calculation results of each batch during training. In this embodiment, the loss function for each batch during training is the average cross-entropy loss of each dual-channel feature sample in the current batch during the training of the hybrid neural network model.

[0116] When calculating the cross-entropy loss function of each dual-channel feature sample in the current batch for training the hybrid neural network model, the true oscillation category of each dual-channel feature sample is converted into a one-hot encoding form: if the true oscillation category of a certain dual-channel feature sample is forced oscillation, then the encoding value of the position corresponding to forced oscillation is 1 and the encoding value of the position corresponding to free oscillation is 0; if the true oscillation category of a certain dual-channel feature sample is free oscillation, then the encoding value of the position corresponding to forced oscillation is 0 and the encoding value of the position corresponding to free oscillation is 1.

[0117] When calculating the cross-entropy loss, based on the one-hot encoding of the true oscillation class of the sample, the predicted probability value corresponding to that true class is selected from the two predicted probability values ​​output by the model. Specifically, if the true class of the sample is forced oscillation (i.e., the one-hot encoding value corresponding to forced oscillation is 1), the predicted probability value of forced oscillation output by the model is taken; if the true class is free oscillation, the predicted probability value of free oscillation output by the model is taken. Then, the negative of the natural logarithm of the selected predicted probability value is calculated. This calculation result is the cross-entropy loss value of the sample for the hybrid neural network model during the current batch training.

[0118] The average cross-entropy loss value for the current batch of training is obtained by summing the cross-entropy loss values ​​of all dual-channel feature samples used in the training of the hybrid neural network model and dividing the sum by the total number of samples in the batch. The goal of model training is to minimize this average cross-entropy loss by iteratively updating the trainable parameters.

[0119] The update process employs a phased training strategy, sequentially training the front-end CNN spatial feature extractor and the back-end LSTM temporal analyzer and classification layer. Both phases use the backpropagation algorithm and a gradient-based optimizer for parameter updates.

[0120] In the first stage, dual-channel feature samples are input into the CNN spatial feature extractor, and the stochastic gradient descent (SGDM) optimizer is used for 8 rounds of training to update all trainable parameters of the CNN spatial feature extractor. After training, the parameters of the CNN spatial feature extractor are fixed.

[0121] In the second stage, the 3D fused feature tensor output by the CNN spatial feature extractor is reshaped and input into the LSTM temporal analyzer. The Adam optimizer is used for 15 epochs of training, updating all trainable parameters of the LSTM temporal analyzer and the fully connected layers in the classification layer. The batch size for both stages is 32, and training termination is based on loss function convergence.

[0122] The model parameters that need to be updated include the kernel weights and biases of all convolutional blocks in the CNN spatial feature extractor in the first stage, the scaling and translation parameters of the batch normalization layer, and the weights and biases of the temporary fully connected layer at the end of the CNN used for the first stage training; and the weight matrices and bias vectors of the forget gate, input gate, candidate memory unit, and output gate in the LSTM temporal analyzer in the second stage, as well as the weight matrix and bias vectors of the fully connected layer in the final classification layer.

[0123] Therefore, in this embodiment, a hybrid neural network model is trained using a training set to obtain a trained hybrid neural network model, which serves as a low-frequency oscillation category recognition model.

[0124] Step 3: Obtain the low-frequency oscillation signal sequence to be identified in the power system. Based on the same process as in Step 1, construct the time-domain energy map and phase plane trajectory map of the low-frequency oscillation signal sequence to be identified in the power system, forming the current dual-channel feature sample.

[0125] The current dual-channel feature sample is input into the low-frequency oscillation category recognition model, and the low-frequency oscillation category recognition model is trained to output the oscillation category recognition result corresponding to the current dual-channel feature sample.

[0126] To verify the effectiveness of the method proposed in this embodiment, an improved IEEE 4-unit two-zone system simulation model with a grid-type wind turbine was built based on the DigSILENT simulation platform. The topology of the improved IEEE 4-unit two-zone system simulation model is as follows: Figure 4 As shown, it includes WTG1 grid-connected doubly-fed wind turbine generators, synchronous generators G1, G2, G3, and G4. Figure 4Numbers 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, and 11 are all buses in the improved IEEE 4-machine two-zone system. Among them, the WTG1 grid-type doubly-fed wind turbine is connected to bus 12; synchronous generator G1 is connected to bus 1, synchronous generator G2 is connected to bus 2, synchronous generator G3 is connected to bus 3, and synchronous generator G4 is connected to bus 4.

[0127] Forced oscillations were induced by applying periodic sinusoidal disturbances to different control loops of GFM-DFIG using the parameter scanning method. Free oscillations were induced by adjusting the PSS gain and virtual damping coefficient and setting a three-phase short-circuit fault. A total of 1100 samples were generated and divided into training and test sets at a ratio of 4:1.

[0128] The model is trained using the training set and finally evaluated using the test set. Comparison results with methods such as SVM, Random Forest, single CNN, and single LSTM are shown below. Figure 5 As shown, Figure 5 The CNN-LST curve in this embodiment shows the accuracy curve. Figure 5 As can be seen, the method in this embodiment achieves the highest identification accuracy under different noise conditions, and still maintains an accuracy of 96.4% under 20dB strong noise conditions, demonstrating excellent noise resistance robustness.

[0129] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. These embodiments are merely descriptions of preferred embodiments and are not intended to limit the scope or concept of the invention. The specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. Such combinations, as long as they do not violate the spirit of the present invention, should also be considered as part of this disclosure. To avoid unnecessary repetition, the present invention will not further describe the various possible combinations.

[0130] This invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of this invention and without departing from the design idea of ​​this invention, all modifications and improvements made by those skilled in the art to the technical solutions of this invention should fall within the protection scope of this invention. The technical content for which protection is sought in this invention has been fully described in the claims.

Claims

1. A method for identifying low-frequency oscillation categories in power systems based on multi-physical feature fusion and deep learning, characterized in that, The process is as follows: Step 1: Obtain low-frequency oscillation signal sequences of the power system over multiple time periods. The true oscillation category of the low-frequency oscillation signal sequence for each time period is known, and the true oscillation category is one of forced oscillation or free oscillation. Establish the time-domain energy spectrum and phase plane trajectory map of the low-frequency oscillation signal sequence for each time period. Use the time-domain energy spectrum and phase plane trajectory map of the low-frequency oscillation signal sequence for each time period as dual-channel feature samples, and form a training set from the dual-channel feature samples of each time period. Step 2: Generate a hybrid neural network model and train the hybrid neural network model using the dual-channel fused feature map samples in the training set obtained in Step 1. During training, a hybrid neural network model is used to obtain the corresponding dual-channel fusion feature map based on each dual-channel fusion feature map sample. Then, a three-dimensional fusion feature tensor is extracted from the dual-channel fusion feature map. Subsequently, the dual-channel fusion feature map is reshaped into a feature sequence, and the dynamic evolution law of the feature sequence is explored. Finally, linear transformation and activation are performed based on the dynamic evolution results to obtain the final predicted oscillation category of each dual-channel fusion feature map sample. During training, the parameters of the hybrid neural network model are updated based on the calculation results of the loss function, and the trained hybrid neural network model is used as a low-frequency oscillation category recognition model. Step 3: Obtain the time-domain energy spectrum and phase plane trajectory diagram of the low-frequency oscillation signal sequence to be identified in the power system, and form the current dual-channel feature sample; The current dual-channel feature sample is input into the low-frequency oscillation category recognition model, and the low-frequency oscillation category recognition model outputs the oscillation category recognition result corresponding to the current dual-channel feature sample.

2. The method for identifying low-frequency oscillation categories in power systems based on multi-physical feature fusion and deep learning according to claim 1, characterized in that, In step 1, wavelet transform is performed on the low-frequency oscillation signal sequence for each time period to obtain the time-domain energy map corresponding to the low-frequency oscillation signal sequence for each time period.

3. The method for identifying low-frequency oscillation categories in power systems based on multi-physical feature fusion and deep learning according to claim 1, characterized in that, In step 1, phase plane analysis is performed on the low-frequency oscillation signal sequence for each time period to obtain the phase plane trajectory diagram of the low-frequency oscillation signal sequence for each time period.

4. The method for identifying low-frequency oscillation categories in power systems based on multi-physical feature fusion and deep learning according to claim 1, characterized in that, In step 2, the hybrid neural network model includes a dual-channel input layer, a CNN spatial feature extractor, a feature reshaping layer, an LSTM temporal analyzer, and a classification layer; During training, a dual-channel fusion feature map is obtained based on each dual-channel fusion feature map sample through a dual-channel input layer; a three-dimensional fusion feature tensor is extracted from the dual-channel fusion feature map through a CNN spatial feature extractor; the dual-channel fusion feature map is reshaped into a feature sequence through a feature reshaping layer; the dynamic evolution pattern of the feature sequence is mined through an LSTM time series analyzer; finally, a classification layer performs linear transformation and activation based on the dynamic evolution results to obtain the final predicted oscillation category of each dual-channel fusion feature map sample.

5. The method for identifying low-frequency oscillation categories in power systems based on multi-physical feature fusion and deep learning according to claim 4, characterized in that, The CNN spatial feature extractor includes three cascaded convolutional blocks. Each convolutional block includes a learnable convolutional kernel, batch normalization, ReLU activation function, and max pooling. The first convolutional block takes each dual-channel fused feature map as input, the second convolutional block takes the output feature map obtained from the first convolutional block as input, and the third convolutional block takes the output feature map obtained from the second convolutional block as input. In each convolutional block, a sliding window operation is performed on the input feature map using a learnable filter to extract local spatial features. Then, batch normalization, ReLU activation, and max pooling are applied to the local spatial features to obtain the output feature map. Finally, the three-dimensional fused feature tensor is obtained by the third convolutional block.

6. The method for identifying low-frequency oscillation categories in power systems based on multi-physical feature fusion and deep learning according to claim 4, characterized in that, The feature reshaping layer performs serialized slicing and flattening of the three-dimensional fused feature tensor along the time axis to obtain the feature sequence.

7. The method for identifying low-frequency oscillation categories in power systems based on multi-physical feature fusion and deep learning according to claim 4, characterized in that, The LSTM time series analyzer includes multiple cascaded LSTM layers. The feature sequence is input into the first LSTM layer, and the dynamic evolution pattern is mined step by step through multiple LSTM layers to obtain the hidden state vector of the last time step as the dynamic evolution result of the feature sequence.

8. The method for identifying low-frequency oscillation categories in power systems based on multi-physical feature fusion and deep learning according to claim 4, characterized in that, The classification layer includes a fully connected layer and a Softmax activation function. The two-dimensional vector is obtained by linear transformation of the dynamic evolution result of the classification layer. Then, the two-dimensional vector is activated by the Softmax activation function to obtain the two-dimensional probability distribution of the predicted oscillation category of each dual-channel feature sample. The predicted oscillation category with the larger probability value is taken as the final predicted oscillation category of each dual-channel feature sample.

9. The method for identifying low-frequency oscillation categories in power systems based on multi-physical feature fusion and deep learning according to claim 4, characterized in that, For the parameters of the hybrid neural network model, the update process adopts a staged training strategy; In the first stage, dual-channel feature samples are input into the CNN spatial feature extractor, and all trainable parameters of the CNN spatial feature extractor are updated using stochastic gradient descent. In the second stage, the 3D fusion feature tensor output by the CNN spatial feature extractor is reshaped and then input into the LSTM temporal analyzer. The Adam optimizer is used to update all trainable parameters of the LSTM temporal analyzer and the fully connected layers in the classification layer.

10. The method for identifying low-frequency oscillation categories in power systems based on multi-physical feature fusion and deep learning according to claim 1, characterized in that, In step 2, the loss function during training is calculated as the average cross-entropy loss of each dual-channel feature sample during the training of the hybrid neural network model.

Images