Complex power quality disturbance identification method based on OMD and two-dimensional image coding

Abstract

The invention provides a complex power quality disturbance identification method based on OMD and two-dimensional image coding. The complex power quality disturbance identification method comprises the steps that 1, power quality disturbance signals are collected; 2, carrying out sparse transformation on the power quality disturbance signal by adopting a compressed sensing technology, and compressing the signal dimension; 3, performing optimal mode decomposition on the compressed signal, and extracting a multi-mode component; step 4, converting the optimal modal component into a Gramb angle field image and a recurrence plot, and generating a two-dimensional feature enhanced image after feature decomposition and fusion based on correlation driving; and step 5, inputting the fused two-dimensional image into the improved convolutional neural network for identification to obtain an identification result. According to the method, one-dimensional signals are converted into two-dimensional images, feature extraction is added, the potential of multi-model fusion is further mined, and the method is suitable for larger-scale complex power quality problem classification application.

Description

Technical Field

[0001] The invention relates to the technical field of power quality disturbance identification, and in particular to a complex power quality disturbance identification method based on OMD and two-dimensional image coding. Technical Background

[0002] In recent years, new energy technologies have developed rapidly, and a large number of new energy distributed power sources have been connected to the power grid, prompting the power system to gradually evolve from the traditional decentralized radial distribution network structure to a multi-terminal network structure with multiple power nodes. This transformation has brought about a series of power quality problems such as power harmonics, three-phase voltage imbalance, and voltage fluctuations and flickers. In addition, the widespread application of power electronic equipment in power systems has further aggravated the severity of power quality problems. In order to ensure the stable operation of the power system and the reliability of power quality, it is necessary to pay comprehensive attention to and solve power quality disturbance-related problems in all aspects of the power system.

[0003] The patent with application publication number CN117932402A discloses a power quality disturbance classification method and system based on one-dimensional convolutional neural network multi-model fusion. This method proposes a parameter-based learnable result fusion model by combining time domain, frequency domain and time-frequency domain features, thereby significantly improving the classification performance of disturbance signals. Specifically, a time domain one-dimensional convolutional neural network model based on the original signal, a frequency domain one-dimensional convolutional neural network model based on Fourier transform, and a time-frequency domain one-dimensional convolutional neural network model based on wavelet transform are constructed, and multi-model fusion is realized through BP neural network. The patent with application publication number CN118551285A discloses a power quality disturbance intelligent identification method based on a multimodal feature fusion network. This method fully mines the effective information in the disturbance signal through feature fusion and classification technology, thereby significantly improving the recognition accuracy of various power quality disturbance problems. Specifically, the invention proposes a model architecture that combines time-domain structural feature extraction and multimodal feature fusion, simplifies the manual design of complex disturbance patterns in traditional methods, and comprehensively optimizes the recognition effect through model training, verification and testing.

[0004] In summary, the feature selection of the above methods is still limited to the time domain, frequency domain and time-frequency domain features, which may not fully cover all the effective information in the power disturbance signal, and there is potential information loss. Summary of the invention

[0005] The purpose of the present invention is to provide a complex power quality disturbance identification method based on OMD and two-dimensional image coding, in view of the fact that feature selection is limited to one-dimensional features and there is a potential information loss problem in traditional power quality disturbance identification, which converts one-dimensional signals into two-dimensional images, increases feature extraction, and further explores the potential of multi-model fusion, which is suitable for larger-scale complex power quality problem classification applications.

[0006] To achieve the above-mentioned invention purpose, the method of the present invention includes the following steps:

[0007] Step 1: Collect power quality disturbance signals;

[0008] Step 2: Use compressive sensing technology to perform sparse transformation on the power quality disturbance signals to obtain compressed signals;

[0009] Step 3: Perform optimal mode decomposition on the compressed signals to extract the optimal modal components;

[0010] Step 4: Convert the optimal modal components into Gramian angular field images and recurrence plots, and generate two-dimensional feature enhanced images through fusion based on correlation-driven feature decomposition;

[0011] Step 5: Input the fused two-dimensional feature enhanced images into an improved convolutional neural network for recognition to obtain the power quality disturbance recognition results.

[0012] Further, in Step 1, collect the power quality disturbance signal x(t), and sample it at a fixed sampling frequency f s to obtain the discrete-time sequence signal x(n), where n = 0, 1, 2,..., N - 1; perform denoising processing on the discrete-time sequence signal x(n), and normalize the signal amplitude range to [-1, 1].

[0013] Further, Step 2 includes:

[0014] First, use the orthogonal matching pursuit method to perform sparse representation on the denoised discrete-time sequence signal x(n) under the given sensing matrix φ, and obtain the sparse coefficient z; then, train the sparse dictionary through the K-singular value decomposition method, and optimize the sparse dictionary D and the sparse coefficient matrix Z at the same time to complete the construction of the sparse dictionary; finally, based on the compressive sensing theory, project the measurement signal into a low-dimensional space to generate the compressed signal y(t).

[0015] Further, Step 3 specifically includes:

[0016] Step 3-1: Apply empirical mode decomposition to the compressed signal y(t) to decompose it into multiple intrinsic mode functions;

[0017] Step 3-2: Calculate the total energy of the original signal and the total energy of all intrinsic mode function components, obtain the ratio of the total energy of the intrinsic mode functions to the total energy of the original signal, and define it as the energy ratio ξ;

[0018] Step 3-3: Set the range of the number of mode decompositions K, for different K values, repeat Steps 3-1 to 3-2, calculate the corresponding energy ratio ξ, and determine the number of modes K that makes the energy ratio ξ the largest optimal ;

[0019] Step 3-4: Use the number of modes K optimal Perform optimal mode decomposition on the compressed signal y(t) to obtain the optimal mode components IMF p (t).

[0020] Furthermore, step 4 specifically includes:

[0021] Step 4-1: Convert the one-dimensional time series into a recurrence plot through the distance between vectors in the high-dimensional phase space;

[0022] Step 4-2: Convert the one-dimensional time series into a Gramian angular field image based on polar coordinate encoding and the Gram matrix;

[0023] Step 4-3: Extract the shallow features of the recurrence plot and the Gramian angular field image through a shared encoder, and use a basic feature encoder and a detail feature encoder to extract low-frequency basic information and high-frequency detail information;

[0024] Step 4-4: In the fusion layer, fuse the low-frequency basic information and high-frequency detail information of the recurrence plot and the Gramian angular field image respectively;

[0025] Step 4-5: Generate a two-dimensional feature enhanced image through a decoder.

[0026] Furthermore, in step 5, the improved convolutional neural network includes an input layer, a hierarchical DenseNet feature extraction module, a feature enhancement SimAM module, a global pooling layer, a fully connected layer, a Softmax layer, and an output layer;

[0027] Use the obtained two-dimensional feature enhanced image as the source image and input it into the output layer of the improved convolutional neural network. Among them, the hierarchical DenseNet feature extraction module is the core module of the improved convolutional neural network. The hierarchical DenseNet feature extraction module contains 5 layers of two-dimensional convolution, and each layer is followed by a feature enhancement SimAM module to perform dynamic weight allocation on the input features to highlight key features; the output of each layer is directly passed to all subsequent layers to form dense connections; after the last layer of convolution, use the global pooling layer to reduce the features to a fixed size; input the output features of the hierarchical DenseNet feature extraction module into the fully connected layer for feature fusion, and the Softmax layer maps the output of the fully connected layer to a probability distribution to output the final power quality disturbance recognition result.

[0028] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0029] 1. The present invention abandons the traditional classification method that relies on manual feature design and completely adopts a data-driven approach for feature learning and classification, thereby reducing manual intervention and improving the generalization ability of the model. Based on the automatic feature extraction framework of deep learning, the present invention can be flexibly extended to other types of perturbations or data patterns and is applicable to more complex power quality perturbation detection tasks.

[0030] 2. The present invention combines compressive sensing and optimized matching pursuit techniques to generate a high-quality sparse feature matrix and inputs it into a classifier for perturbation type identification. By combining the optimization algorithm and the feature extraction framework, the present invention can process a large amount of complex data in a short time, ensuring that during the operation of the power system, it can detect and identify perturbation types in real time, thereby realizing the dynamic monitoring of power quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The present invention will be further described below in conjunction with the drawings and embodiments:

[0032] Figure 1 is a flowchart of a complex power quality perturbation identification method based on OMD and two-dimensional image coding according to an embodiment of the present invention;

[0033] Figure 2 is a schematic diagram of the network structure of an improved convolutional neural network proposed by the present invention;

[0034] Figure 3 is a schematic diagram of an improved hierarchical dense network layer based on the attention mechanism proposed by the present invention;

[0035] Figure 4 is a graph showing the relationship between accuracy and different sparse dictionary sizes in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Apparently, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0037] The embodiments of the present invention provide a complex power quality perturbation identification method based on OMD and two-dimensional image coding, as Figure 1 shown, including the following steps:

[0038] Step 1: Collect power quality perturbation signals;

[0039] Step 2: Use compressive sensing technology to perform sparse transformation on the power quality disturbance signal to obtain a compressed signal;

[0040] Step 3: Perform optimal mode decomposition on the compressed signal to extract the optimal mode components;

[0041] Step 4: Convert the optimal mode components into Gramian angular field images and recurrence plots, and generate a two-dimensional feature enhanced image through fusion based on correlation-driven feature decomposition;

[0042] Step 5: Input the fused two-dimensional feature enhanced image into an improved convolutional neural network for recognition to obtain the recognition result.

[0043] In Step 1, collect the power quality disturbance signal x(t), and sample it at a fixed sampling frequency f s to obtain a discrete-time series signal x(n), where n = 0, 1, 2,..., N - 1. Denoise the sampled signal to eliminate noise interference, ensure the signal is clear, and normalize the signal amplitude range to [-1, 1].

[0044] In Step 2, use compressive sensing technology to perform sparse transformation on the power quality disturbance signal to compress the signal dimension.

[0045] The specific steps are as follows:

[0046] Step 2-1: Use orthogonal matching pursuit to perform sparse representation on the input signal x under the given sensing matrix φ, and find the sparse representation z that minimizes the measurement error. The optimization objective is as follows:

[0047]

[0048] In the formula: z t is the sparse coefficient vector obtained in the t-th iteration; x is the input power quality disturbance signal; φ is the sensing matrix used to project the signal into a low-dimensional space; z is the sparse coefficient vector; ||·||2 is the two-norm of the vector, which is the length or size of the vector in the Euclidean space.

[0049] Step 2-2: Use the K-singular value decomposition algorithm to train the sparse dictionary, and simultaneously optimize the sparse dictionary D and the sparse coefficient matrix Z to minimize the error between the original signal matrix X and its sparse reconstruction DZ. The formula is as follows:

[0050]

[0051] In the formula: X is the training signal matrix, with dimension N×K, and each column is a training signal sample; D is the sparse dictionary matrix, with dimension N×L, and each column is a dictionary atom; Z is the sparse coefficient matrix, with dimension L×K, and each column is the sparse representation of a signal; ||·||F is the Frobenius norm, which is used to measure the reconstruction error of the matrix X - DZ.

[0052] Step 2 - 3: After the construction of the sparse dictionary is completed, project the sparse signal obtained in Step 2 - 1 into a low - dimensional space for compressed storage. Measure the power quality disturbance signal x through the sensing matrix φ to generate the compressed measurement signal y.

[0053] To further compress the signal storage, assume that the signal x is a sparse linear combination of the sparse dictionary D, that is:

[0054] x = Dz (3)

[0055] Substitute x into the measurement formula to get:

[0056] y = φDz (4)

[0057] Where: y is the compressed measurement signal; x is the power quality disturbance signal.

[0058] Thus, this step completes the conversion from the sparse signal to the compressed measurement signal.

[0059] Step 3: Perform the best - mode decomposition on the compressed signal to extract multi - modal components.

[0060] Step 3 - 1: Apply empirical mode decomposition to y(t) and decompose it into multiple intrinsic mode functions:

[0061]

[0062] Where: K is the number of decomposed modes, and p represents the p - th intrinsic mode function.

[0063] Step 3 - 2: First, calculate the total energy of the original signal and the total energy of all intrinsic mode function components:

[0064]

[0065] Where: S is the number of sampling points, and y(j) is the amplitude of the signal at the sampling point;

[0066] Define the energy ratio ξ according to the ratio of the total energy of the intrinsic mode function to the total energy of the original signal:

[0067]

[0068] Step 3 - 3: Set the range of the number of mode decompositions K. For different K values, repeat Steps 3 - 1 to 3 - 2, calculate the corresponding energy ratio ξ, and determine the number of modes K that maximizes the energy ratio ξ optimal ;

[0069] Step 3-4: Use the optimal number of modes K optimal Perform optimal mode decomposition on the input signal to obtain the optimal mode components IMF p (t);

[0070] Step 4: Convert the optimal mode component I into a Gram angle field image and a recurrence plot, and generate a two-dimensional feature enhanced image through fusion based on correlation-driven feature decomposition. Step 4 includes:

[0071] Step 4-1: Convert the one-dimensional time series into a recurrence plot through the distance between vectors in the high-dimensional phase space.

[0072] First, reconstruct the one-dimensional time series I = {i1, i2..., i n} into a phase space trajectory matrix through the delay time r and the embedding dimension d:

[0073]

[0074] where: r is the delay time; d is the embedding dimension.

[0075] Calculate the recurrence value matrix according to the Euclidean distance between vectors:

[0076] RP(i, j) = Heaviside(ε - ||I i - I j ||) (9)

[0077] where: ε is the set distance threshold; Heaviside is the step function.

[0078] Step 4-2: Convert the one-dimensional time series into a Gram angle field image based on polar coordinate encoding and the Gram matrix.

[0079] First, map the one-dimensional time series values to the polar coordinate angle φ, and then define a two-dimensional image using the inner product of the Gram matrix:

[0080]

[0081] Step 4-3: Extract the shallow features of the recurrence plot and the Gram angle field image through a shared encoder, and use a basic feature encoder and a detail feature encoder to extract low-frequency basic information and high-frequency detail information;

[0082] Step 4-4: In the fusion layer, fuse the basic features and detail features of the recurrence plot and the Gram angle field image respectively;

[0083] Step 4-5: Generate a two-dimensional feature enhanced image through a decoder.

[0084] Step 5: Input the fused two-dimensional feature enhanced image into an improved convolutional neural network for recognition to obtain the recognition result. The improved convolutional neural network (as shown in Figure 2 ) includes an input layer, a hierarchical DenseNet feature extraction module, a feature enhancement SimAM module, a global pooling layer, a fully connected layer, a Softmax layer, and an output layer. The perturbation classification process is as follows:

[0085] Step 5-1: Use the obtained two-dimensional feature enhanced image as the source image and input it into the improved convolutional neural network. Among them, the hierarchical DenseNet feature extraction module is the core module of the improved convolutional neural network. The multi-layer convolutional network contains 5 layers of two-dimensional convolution, and each layer is followed by a SimAM module to perform dynamic weight allocation on the input features and highlight key features.

[0086] Specifically, the multi-layer convolutional network contains 5 layers of two-dimensional convolution. The convolution kernel size of each layer is 5×5. The number of feature channels in the first layer is 32, and the rest are 16. The SimAM module is integrated after each convolutional layer to dynamically adjust the output feature map. First, calculate the energy value of each neuron in each feature map:

[0087]

[0088] where: t is the target neuron value; β is the feature map mean; σ is the standard deviation of the feature map; α is a fixed constant with a value of 10 -4 .

[0089] Then calculate the weight of the neuron according to the energy value, apply the weight to the feature map, and adjust the feature through point-by-point multiplication:

[0090] X adjusted =σ(ω t )⊙X (12)

[0091] where: X is the input feature map; X adjusted is the adjusted feature map; ω t is the neuron weight; σ is the Sigmoid activation function used to normalize the weight.

[0092] Through the SimAM module, the key features of complex power disturbance signals are highlighted, and redundant or noise information is suppressed;

[0093] Step 5-2: The output of each layer is directly passed to all subsequent layers to form a dense connection (as shown in Figure 3 ). Specifically, the Mish activation function is applied after each layer of convolution to enhance the nonlinear expression ability of the model;

[0094] Step 5-3: After the last layer of convolution, use a global pooling layer to reduce the dimensionality of the features to a fixed size. Then, input the output features of the hierarchical DenseNet feature extraction module into a fully connected layer for feature fusion;

[0095] Step 5-4: The Softmax layer maps the output of the fully connected layer to a probability distribution and outputs the final power quality disturbance recognition result.

[0096] Example:

[0097] To verify the classification effect of the improved convolutional neural network model, a series of experiments are carried out under different noise interference levels. Specifically, in this example, 28 power quality disturbance models are established according to the IEEE1159 standard, including 9 single disturbances such as normal signal, voltage sag, voltage swell, voltage interruption, harmonics, voltage transient oscillation, voltage flicker, voltage notch, voltage spike, and 19 complex disturbance signals mixed by these single disturbances. As shown in Table 1, there are 28 different power quality disturbance signals. In the experiment, 28 power quality disturbance signals are simulated and generated in MATLAB, with a fundamental frequency of 50Hz, a signal sampling period of 200ms, a sampling frequency of 6.4KHz, 1280 sampling points for each power quality disturbance, and divided into a training set, a validation set, and a test set according to a ratio of 6:2:2. Set the training iteration times of the convolutional attention neural network to 50 and the initial learning rate to 0.001.

[0098] Table 1 28 different power quality disturbance signals

[0099] Type Power quality disturbance signal Type Power quality disturbance signal N1 Normal N15 Rise + Transient N2 Rise N16 Fall + Transient N3 Fall N17 Notch + Transient N4 Transient N18 Spike + Transient N5 Interruption N19 Fall + Flicker N6 Flicker N20 Fall + Flicker + Harmonic N7 Harmonic N21 Rise + Flicker + Harmonic N8 Spike N22 Interruption + Flicker + Harmonic N9 Notch N23 Fall + Transient + Harmonic N10 Rise + Harmonic N24 Rise + Transient + Harmonic N11 Fall + Harmonic N25 Interruption + Transient + Harmonic N12 Flicker + Harmonic N26 Flicker + Transient + Harmonic N13 Interruption + Harmonic N27 Rise + Transient + Spike N14 Transient + Harmonic N28 Rise + Transient + Notch

[0100] In order to reduce the data storage volume and at the same time recover the actual power quality disturbance signal from the compressed form, it is necessary to determine the sparse basis of the power quality disturbance signal. Since the time-domain signal of the power quality disturbance signal is not sparse, these signals are transformed into the sparse domain using orthogonal matching pursuit. Figure 4 It shows the relationship between the accuracy and the change of the sparse dictionary size M from 100 to 300; it is confirmed that M = 300. Therefore, for analysis purposes, the sparse dictionary dimension is fixed at (300×1200).

[0101] To verify the effectiveness of the present invention under different power quality disturbance signals, the classification accuracy of different power quality disturbance signals under various noise conditions is shown in Table 2. It can be clearly seen from Table 2 that the recognition accuracy of the present invention is higher than 99%. When there is no noise interference, the classification performance of the present invention is generally significantly better than that under noise interference. The classification accuracy of the model for a single power quality disturbance signal is significantly higher than that of the combined power quality disturbance signal, which indicates that the characteristics of a single power quality disturbance signal are relatively simple, while the characteristics of the combined power quality disturbance signal affect each other and are more complex. And under the condition that the noise is 20 dB, the minimum accuracy rate of the proposed improved convolutional neural network on N26 is 98.23%. The main reason is that when transient signals are superimposed simultaneously, the flicker is more concealed and the characteristics are more difficult to identify. Nevertheless, the accuracy rate of the present invention is still higher than 98.2%, which proves that the improved convolutional neural network still has excellent feature extraction and disturbance classification performance in complex power quality disturbance detection, and at the same time has a satisfactory anti-noise ability.

[0102] Table 2 Classification results of different power quality disturbance signals under different noises

[0103]

[0104]

[0105] To further evaluate the effects and advantages of the present invention, we compared its experimental results with those of some advanced methods, including traditional power quality disturbance classification techniques (based on artificial feature extraction) and deep learning-based power quality disturbance classification techniques. Traditional methods such as ACMP-SVM, ICEEMDAN+KNN, and TQMT, and deep learning methods including CNN+GRU, CNN+LSTM, DCNN, 1D-ICNN, and EITD+GSCN. Under different noise conditions, the power quality disturbance classification results of various methods are shown in Table 3. As can be seen from the table, the classification accuracy of the artificial feature extraction method is lower than that of the deep learning-based automatic classification method, indicating that the artificial method may cause the loss of some feature information, while the deep learning method can effectively avoid this problem and fully explore the deep features of the power quality disturbance signal. It should be noted that the experimental dataset of the present invention contains 28 power quality disturbance categories, which is the largest. Compared with the experimental datasets of other comparison methods, the power quality disturbance categories in the datasets of other methods are fewer, and these categories are also included in the experimental data of the present invention. Although 1D-ICNN, DCNN, and EITD+GSCN achieved accuracies of 99.36%, 93.10%, and 98.83% respectively, these methods only considered fewer power quality disturbance categories. Our improved convolutional neural network method considered up to 28 power quality disturbance signals and achieved an accuracy of 99.39% at a signal-to-noise ratio of 20 dB. This shows that the improved convolutional neural network method performs better in terms of accuracy and anti-noise ability and is more suitable for application in complex power quality disturbance signal detection tasks.

[0106] Table 3 Classification results of different methods for different power quality disturbance signals

[0107]

[0108]

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications or equivalent replacements can still be made to the specific embodiments of the present invention, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention shall be covered by the protection scope of the claims of the present invention.

Claims

1. A complex power quality disturbance recognition method based on OMD and two-dimensional image coding, characterized in that It includes the following steps: Step 1: Collect power quality disturbance signals; Step 2: Use compressive sensing technology to perform sparse transformation on the power quality disturbance signals to obtain compressed signals; Step 3: Perform optimal mode decomposition on the compressed signals to extract the optimal mode components; Step 4: Convert the optimal mode components into Gram angular field images and recurrence plots, and generate two-dimensional feature enhanced images through correlation-driven feature decomposition fusion; Step 5: Input the fused two-dimensional feature enhanced images into an improved convolutional neural network for recognition to obtain the power quality disturbance recognition results.

2. The complex power quality disturbance recognition method based on OMD and two-dimensional image coding according to claim 1, characterized in that: In step 1, the power quality disturbance signal x(t) is collected and sampled at a fixed sampling frequency f s to obtain the discrete-time series signal x(n), where n = 0, 1, 2, ..., N-1; the discrete-time series signal x(n) is denoised and the signal amplitude range is normalized to [-1, 1].

3. The complex power quality disturbance recognition method based on OMD and two-dimensional image coding according to claim 1, characterized in that: Step 2 includes: First, use the orthogonal matching pursuit method to perform sparse representation on the denoised discrete time series signal x(n) under the given sensing matrix φ and obtain the sparse coefficient z; then, train the sparse dictionary through the K-singular value decomposition method, and optimize the sparse dictionary D and the sparse coefficient matrix Z at the same time to complete the construction of the sparse dictionary; finally, based on the compressive sensing theory, project the measurement signal into a low-dimensional space to generate the compressed signal y(t).

4. The complex power quality disturbance recognition method based on OMD and two-dimensional image coding according to claim 1, characterized in that: Step 3 specifically includes: Step 3-1: Apply empirical mode decomposition to the compressed signal y(t) and decompose it into multiple intrinsic mode functions; Step 3-2: Calculate the total energy of the original signal and the total energy of all intrinsic mode function components, obtain the ratio of the total energy of the intrinsic mode functions to the total energy of the original signal, and define it as the energy ratio ξ; Step 3-3: Set the range of the number of modal decompositions K. For different values of K, repeat Steps 3-1 to 3-2, calculate the corresponding energy ratio ξ, and determine the number of modes K that maximizes the energy ratio ξ optimal ; Step 3-4: Use the number of modes K optimal Perform optimal mode decomposition on the compressed signal y(t) to obtain the optimal mode components IMF p (t).

5. The complex power quality disturbance recognition method based on OMD and two-dimensional image coding according to claim 1, characterized in that: Step 4 specifically includes: Step 4-1: Convert the one-dimensional time series into a recurrence plot through the distance between vectors in the high-dimensional phase space; Step 4-2: Convert the one-dimensional time series into a Gram angular field image based on polar coordinate encoding and the Gram matrix; Step 4-3: Extract the shallow features of the recurrence plot and the Gram angular field image through a shared encoder, and use a basic feature encoder and a detail feature encoder to extract low-frequency basic information and high-frequency detail information; Step 4-4: In the fusion layer, fuse the low-frequency basic information and high-frequency detail information of the recurrence plot and the Gram angular field image respectively; Step 4-5: Generate a two-dimensional feature enhanced image through a decoder.

6. The complex power quality disturbance identification method based on OMD and two-dimensional image coding according to claim 1, characterized in that: In Step 5, the improved convolutional neural network includes an input layer, a hierarchical DenseNet feature extraction module, a feature enhanced SimAM module, a global pooling layer, a fully connected layer, a Softmax layer, and an output layer; Take the obtained two-dimensional feature enhanced image as the source image and input it into the output layer of the improved convolutional neural network. Among them, the hierarchical DenseNet feature extraction module is the core module of the improved convolutional neural network. The hierarchical DenseNet feature extraction module contains 5 layers of two-dimensional convolution, and each layer is followed by a feature enhanced SimAM module to perform dynamic weight allocation on the input features to highlight key features; the output of each layer is directly passed to all subsequent layers to form dense connections; after the last layer of convolution, use the global pooling layer to reduce the features to a fixed size; input the output features of the hierarchical DenseNet feature extraction module into the fully connected layer for feature fusion, and the Softmax layer maps the output of the fully connected layer into a probability distribution to output the final power quality disturbance recognition result.

Images