A power quality composite disturbance identification method, system, device and medium

Abstract

The application discloses a power quality composite disturbance identification method, system, device and medium, and belongs to the technical field of intelligent monitoring and signal processing of power systems, which comprises the following steps: obtaining a one-dimensional time domain feature tensor after standardizing original collected signals, and outputting a time-frequency dual-domain feature tensor through dual-domain subspace feature enhancement; performing time-frequency dual interaction guided fusion on the one-dimensional time domain feature tensor and the time-frequency dual-domain feature tensor to obtain a soft and hard interaction fusion feature tensor; performing multi-granularity semantic mining and aggregation on the soft and hard interaction fusion feature tensor to generate a multi-granularity information high-dimensional robust feature tensor; and performing probability mapping on the multi-granularity information high-dimensional robust feature tensor through a full connection classification module to generate a power quality composite disturbance type identification result. The application can realize detection and classification of single disturbance and composite disturbance under the complex electromagnetic environment and strong background noise interference of a distribution network.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and signal processing technology for power systems, specifically to a method, system, device, and medium for identifying complex power quality disturbances. Background Technology

[0002] Current industrial signal monitoring technology still faces significant bottlenecks in dealing with strong noise interference, multi-source feature fusion, and refined identification, making it difficult to balance signal purity and decision-making accuracy. Specifically, it suffers from the following shortcomings: While existing methods employ denoising techniques such as Singular Value Decomposition (SVD), they are mostly limited to a single scale or a single transform domain. Under strong background noise interference, traditional methods struggle to accurately define the singular value boundaries between signal and noise, resulting in either residual environmental noise or accidental deletion of key physical features during reconstruction, failing to provide high-purity input data for subsequent models.

[0003] Time-domain waveforms and frequency-domain spectra contain complementary physical information, but existing fusion strategies mostly employ simple linear splicing or superposition. This hard fusion lacks dynamic interaction and filtering mechanisms between features, cannot effectively suppress interference from redundant channels, leads to difficulties in semantic alignment between heterogeneous features, and makes it difficult to leverage the synergistic advantages of multi-source information.

[0004] For complex perturbations or weak fault signals, the key discrimination criteria are often hidden in subtle local changes. However, conventional classification decision layers rely excessively on global feature aggregation and lack multi-granular perception capabilities. This coarse feature extraction method can easily lead to micro-details being obscured by macro-information, resulting in insufficient discrimination ability of the model when processing highly similar samples.

[0005] The identification of complex power quality disturbances has evolved from manual features to deep learning. Faced with strong background noise and highly similar disturbances, existing methods have limitations in signal denoising preprocessing, multi-domain fusion, and feature extraction. In preprocessing, improvements such as Singular Value Decomposition (SVD) often focus on determining the order based on difference spectrum or entropy methods, but processing is often limited to a single transform domain, ignoring the sparsity differences of signals in different domains. In strong noise environments, singular value aliasing leads to unclear boundaries, easily introducing noise residue due to retaining too many singular values, or excessively truncating and losing transient features. In multi-domain fusion, two-stream networks typically employ linear splicing or element-wise addition for hard fusion, lacking dynamic interaction between channels and failing to consider the differences in the importance of time-frequency domain information, resulting in redundant interference and difficulty in aligning the semantics of heterogeneous features. In the feature compression stage, conventional global average pooling smooths features in a global macroscopic aggregation manner, masking local high-frequency microscopic discrimination details, making it difficult to improve the model's accuracy in identifying highly similar complex disturbances.

[0006] Of course, many existing papers and patents propose feature extraction methods using a fusion of discrete wavelet transform and hyperbolic S-transform, as well as hybrid signal processing techniques based on Stockwell transform and Hilbert transform, and power quality composite disturbance identification methods based on LightGBM. However, while existing noise reduction methods employ SVD, they are mostly limited to a single scale or transform domain, making it difficult to define singular value boundaries under strong background noise, resulting in residual noise or accidental deletion of electrical features during reconstruction. Time-frequency domain feature fusion is mainly based on linear splicing, lacking dynamic interaction and filtering, and cannot suppress redundant interference, leading to difficulties in semantic alignment of heterogeneous features. Classification decisions rely on global aggregation, and subtle local transient waveform changes are masked by macroscopic statistics, resulting in insufficient ability to identify similar composite disturbances. Summary of the Invention

[0007] In view of the above-mentioned problems, the present invention is proposed.

[0008] Therefore, the present invention aims to solve the problem that strong background noise in the complex electromagnetic environment of power distribution networks masks the characteristics of weak voltage disturbances, resulting in insufficient noise resistance robustness of existing methods.

[0009] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for identifying composite power quality disturbances, comprising, After standardizing the acquired signals from the distribution network's common connection point, a one-dimensional time-domain feature tensor is obtained. This one-dimensional time-domain feature tensor is then enhanced with dual-domain subspace features to output a time-frequency dual-domain feature tensor. The one-dimensional time-domain feature tensor and the time-frequency dual-domain feature tensor are then fused using time-frequency dual-interaction guidance to obtain a hardware-software interactive fusion feature tensor. This hardware-software interactive fusion feature tensor undergoes multi-granularity semantic mining and aggregation to generate a multi-granularity information high-dimensional robust feature tensor. Based on this multi-granularity information high-dimensional robust feature tensor, a probability mapping is performed using a fully connected classification module to generate power quality composite disturbance type identification results.

[0010] As a preferred embodiment of the power quality composite disturbance identification method of the present invention, wherein: the dual-domain subspace feature enhancement includes: The raw one-dimensional voltage data of the common connection point of the power distribution network is standardized and converted into a one-dimensional time-domain feature tensor. A dual-branch parallel enhancement architecture is constructed, which includes a first branch architecture and a second branch architecture; The first branch architecture outputs a time-frequency domain feature tensor from the one-dimensional time-domain feature tensor through time-frequency weighted singular value decomposition and spatial alignment; The second branch architecture constructs the Hankel trajectory tensor from the one-dimensional time-domain feature tensor through a multi-scale sliding window, and obtains the time-domain tensor by performing singular value decomposition. The time-frequency domain feature tensor and the time domain tensor are concatenated along the channel dimension to obtain the time-frequency dual-domain feature tensor.

[0011] As a preferred embodiment of the power quality composite disturbance identification method of the present invention, wherein: the time-frequency weighted singular value decomposition includes: Short-time Fourier transform is performed on the one-dimensional time-domain feature tensor to decompose the non-stationary voltage signal into a two-dimensional time-frequency tensor. The local amplitude energy distribution in the time-frequency plane is calculated from the two-dimensional time-frequency tensor, the perturbation region and noise background are identified, and the time-frequency energy weight tensor is obtained. The weighted time-frequency tensor is obtained by multiplying the two-dimensional time-frequency tensor element by element with the time-frequency energy weight tensor; The time-frequency singular value tensor is decomposed using singular value decomposition, and random noise in the weighted time-frequency tensor is filtered out to output the time-frequency singular value tensor.

[0012] As a preferred embodiment of the power quality composite disturbance identification method of the present invention, the time-frequency dual interactive guided fusion includes: Map the one-dimensional time-domain feature tensor to the query tensor, and map the time-frequency dual-domain feature tensor to the key tensor and the value tensor; Calculate the tensor product of the query tensor and the key tensor, extract the globally weighted feature vector, and expand it to the same spatial dimension as the time-frequency dual-domain feature tensor to obtain the time-domain and frequency-domain feature fusion soft tensor; The one-dimensional time-domain feature tensor is processed by a fully connected layer, dimension reshaping and sigmoid function to obtain a hard attention weight tensor. The hard attention weight tensor is then multiplied element-wise with the time-frequency dual-domain feature tensor to directly filter the effective channels in the frequency domain and obtain the time-frequency domain feature fusion hard tensor. By concatenating the soft tensor of time-domain and frequency-domain feature fusion with the hard tensor of time-domain and frequency-domain feature fusion, a soft-hard interactive fusion feature tensor is obtained.

[0013] As a preferred embodiment of the power quality composite disturbance identification method of the present invention, the multi-granularity semantic mining and aggregation includes: The soft-hard interaction fusion feature tensor is input into the global average pooling layer for processing to obtain the global average pooling tensor and then flattened to obtain the global flattened feature tensor. The soft-hard interaction fusion feature tensor is input into the adaptive pooling layer for processing to obtain the adaptive pooling tensor and then a flattening operation is performed to obtain the adaptive flattened feature tensor. The soft-hard interaction fusion feature tensor is used to obtain a gradient-guided refinement feature tensor by using the gradient-guided detail refinement method, and then a flattening operation is performed to obtain a detail flattening feature tensor. The obtained global flattened feature tensor, adaptive flattened feature tensor, and detailed flattened feature tensor are concatenated along the feature dimension, and a high-dimensional robust feature tensor with multi-granularity information is obtained through multi-granularity aggregation.

[0014] As a preferred embodiment of the power quality composite disturbance identification method of the present invention, the gradient-guided detail refinement method includes: The gradient-guided detail refinement method processes the first and second paths in parallel. In the first path, the soft and hard interaction fusion feature tensor is processed through a deep convolutional layer, and the resulting texture feature tensor is nonlinearly transformed to generate a texture refinement feature tensor. In the second path, the soft and hard interaction fusion feature tensor is processed by the max pooling layer and the average pooling layer respectively and then concatenated to obtain the combined feature tensor. The combined feature tensor is then subjected to gradient sensing to generate the spatial gradient attention weight tensor. The texture refinement feature tensor is concatenated with the spatial gradient attention weight tensor to obtain a weighted texture feature tensor. The weighted texture feature tensor is then convolved and added element-wise with the soft-hard interaction fusion feature tensor to obtain the gradient-guided refinement feature tensor.

[0015] As a preferred embodiment of the power quality composite disturbance identification method of the present invention, the step of probability mapping through a fully connected classification module includes: The multi-granularity information high-dimensional robust feature tensor is mapped to the power quality disturbance category space, and linearly weighted through a fully connected layer to obtain the logistic regression tensor; The logistic regression tensor is subjected to probability normalization using an activation function to obtain a probability distribution tensor; Perform a maximum value indexing operation on the probability distribution tensor, select the category index corresponding to the element with the largest value in the probability distribution as the final judgment basis, and generate the final power quality composite disturbance type identification result.

[0016] Another objective of this invention is to provide a power quality composite disturbance identification system.

[0017] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a power quality composite disturbance identification system, comprising: a dual-domain subspace feature enhancement module, a time-frequency dual interactive guided fusion module, a multi-granularity semantic mining and aggregation module, and a fully connected classification module; The dual-domain subspace feature enhancement module standardizes the collected signals from the common connection point of the distribution network to obtain a one-dimensional time-domain feature tensor, and then outputs the one-dimensional time-domain feature tensor as a time-frequency dual-domain feature tensor through dual-domain subspace feature enhancement. The time-frequency dual-interaction guided fusion module performs time-frequency dual-interaction guided fusion on the one-dimensional time-domain feature tensor and the time-frequency dual-domain feature tensor to obtain the software-hardware interactive fusion feature tensor. The multi-granularity semantic mining and aggregation module performs multi-granularity semantic mining and aggregation on the soft-hard interaction fusion feature tensor to generate a multi-granularity information high-dimensional robust feature tensor. The fully connected classification module generates power quality composite disturbance type identification results by performing probability mapping through fully connected classification based on multi-granularity information high-dimensional robust feature tensors.

[0018] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the power quality composite disturbance identification method.

[0019] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the steps of the power quality composite disturbance identification method.

[0020] The beneficial effects of this invention are as follows: This invention designs a dual-domain subspace feature enhancement module (DDSFEM) and its internal time-frequency weighted singular value decomposition (TFWS) submodule, adopting a dual-branch parallel processing architecture. One branch uses Hankel trajectory tensor combined with SVD decomposition to capture the temporal dynamic evolution trajectory; the other branch uses TFWS to perform low-rank reconstruction of the time-frequency tensor after STFT transformation to filter out random noise. Finally, by splicing, the physical complementarity of time-domain and frequency-domain features is achieved, significantly enhancing the signal feature expression under strong noise background. This invention constructs an energy focusing computation unit (EFU) as the core component of TFWS. By performing element-wise product operation, Gaussian kernel convolution, and Sigmoid activation on the time-frequency tensor, a time-frequency energy weight tensor is generated. The EFU can automatically identify and focus on high-energy perturbation regions and suppress low-energy background noise, providing accurate attention distribution for subsequent weighted enhancement.

[0021] Furthermore, this invention proposes a Time-Frequency Dual Interaction Guided Fusion Module (TFDIGFM) that designs a parallel mechanism of soft and hard dual attention channels. The soft attention fusion channel utilizes tensor operations on queries, keys, and values, along with Softmax normalization, to capture the global correlation between temporal abrupt changes and frequency domain distributions; the hard attention fusion channel directly filters the frequency domain channels through fully connected layers and weights generated by the Sigmoid algorithm; the combination of these two approaches achieves deep interaction and semantic alignment between the original one-dimensional temporal information and the enhanced dual-domain features.

[0022] This invention also develops a multi-granularity semantic mining and aggregation module (MGSMA) that extracts semantic features of different granularities through three branches: global average pooling is used to extract macroscopic semantics, adaptive pooling is used to preserve mesoscopic local structures, and gradient-guided detail refinement (GGDR) is used to preserve microscopic pixel-level information. Finally, the three are concatenated to generate a multi-granularity information high-dimensional robust feature tensor (MIHRF), which solves the problem that single-scale features cannot take into account both global overview and local details.

[0023] This invention also designs a gradient-guided detail refinement submodule: as a micro-branch of MGSMA, it contains a texture feature extraction path and a spatial gradient attention path. The former uses depthwise convolution to extract high-frequency textures, while the latter uses a combination of max pooling and average pooling with large kernel convolution to extract significant edge regions; gradient attention is injected into the feature map through residual connections, achieving refined reconstruction of subtle perturbation details. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 The above is a flowchart of a power quality composite disturbance identification method provided in one embodiment of the present invention.

[0026] Figure 2 This is a dual-domain subspace feature enhancement map of a power quality composite disturbance identification method provided in one embodiment of the present invention.

[0027] Figure 3 The total time-frequency weighted singular value decomposition diagram is provided for a power quality composite disturbance identification method according to an embodiment of the present invention.

[0028] Figure 4 The diagram shows the energy focusing calculation of a power quality composite disturbance identification method provided in one embodiment of the present invention.

[0029] Figure 5 This is a time-frequency dual interactive guided fusion diagram of a power quality composite disturbance identification method provided in one embodiment of the present invention.

[0030] Figure 6 This invention provides a multi-granularity semantic mining and aggregation graph for identifying composite power quality disturbances, as an embodiment of the present invention.

[0031] Figure 7This is a detailed gradient-guided diagram illustrating a power quality composite disturbance identification method according to an embodiment of the present invention.

[0032] Figure 8 This is a fully connected classification diagram of a power quality composite disturbance identification method provided in one embodiment of the present invention. Detailed Implementation

[0033] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0034] Example 1, referring to Figure 1 This is one embodiment of the present invention, which provides a method for identifying composite power quality disturbances, including: S100. After standardizing the collected signals from the common connection point of the distribution network, a one-dimensional time-domain feature tensor is obtained. The one-dimensional time-domain feature tensor is then output as a time-frequency dual-domain feature tensor through dual-domain subspace feature enhancement.

[0035] S200. The one-dimensional time-domain feature tensor and the time-frequency dual-domain feature tensor are fused through time-frequency dual interaction to obtain the software-hardware interactive fusion feature tensor.

[0036] S300: Perform multi-granularity semantic mining and aggregation on the feature tensor of software-hardware interaction fusion to generate a high-dimensional robust feature tensor with multi-granularity information.

[0037] S400: Based on the multi-granularity information high-dimensional robust feature tensor, probability mapping is performed through a fully connected classification module to generate power quality composite disturbance type identification results.

[0038] It should be noted that existing technologies still have the problems of difficulty in fully characterizing the physical characteristics of power quality composite disturbances for single-domain features, difficulty in fusing time-frequency heterogeneous features, and the problem that key micro-transient details in composite disturbances are easily lost in deep network downsampling.

[0039] Therefore, in view of the above-mentioned problems, through steps S100-S400, this invention proposes a power quality composite disturbance identification method based on dual-domain subspace enhancement and multi-granularity feature aggregation. To address the problem that existing technologies for noise reduction are ineffective in dealing with strong background noise interference, this invention designs a dual-domain subspace feature enhancement module, whose output contains a dual-channel feature tensor with rich physical meaning and a high signal-to-noise ratio.

[0040] To address the challenge of heterogeneous feature fusion in existing technologies, this invention designs a time-frequency dual interactive guided fusion module, which achieves precise alignment and complementary fusion of time-frequency features in deep semantics.

[0041] To address the problem of local feature loss caused by complex perturbations that existing technologies cannot effectively solve, this invention designs a multi-granularity semantic mining and aggregation module. Feature information is mined from macroscopic, mesoscopic, and microscopic granularities, and the three feature streams are cascaded to form a high-dimensional, robust semantic embedding, significantly enhancing the classifier's ability to discriminate complex complex perturbations.

[0042] Example 2, refer to Figures 1-8 This is one embodiment of the present invention, which provides a method for identifying composite power quality disturbances, including: This invention integrates non-stationary signal time-frequency analysis, tensor algebra theory, singular value decomposition (SVD) denoising technology, and deep learning multi-granularity attention mechanism. It is mainly used to detect and classify single and compound disturbances in the complex electromagnetic environment and strong background noise interference of power distribution networks. The technical solution of this invention spans multiple interdisciplinary technical fields such as smart grid situational awareness, modern digital signal processing, and artificial intelligence pattern recognition. It should be noted that the present invention presupposes a total of 10 categories of power quality signals to be identified, specifically including: ① normal signal, ② voltage sag, ③ voltage swell, ④ voltage interruption, ⑤ harmonics, ⑥ voltage flicker, ⑦ oscillating transients (6 single disturbances), ⑧ voltage sag + harmonics, ⑨ voltage swell + harmonics, ⑩ oscillating transients + harmonics (3 composite disturbances).

[0043] In this embodiment of the invention, after standardizing the collected signals from the distribution network common connection point in step S100, a one-dimensional time-domain feature tensor is obtained. The one-dimensional time-domain feature tensor is then output as a time-frequency dual-domain feature tensor through dual-domain subspace feature enhancement, including the following steps S101-S103: The original one-dimensional voltage signal of the point of common coupling (PCC) of the distribution network is acquired and standardized to obtain a one-dimensional time-domain feature tensor YF. YF is input to the dual-domain subspace feature enhancement module DDSFEM and outputs a time-frequency dual-domain feature tensor TF.

[0044] S101. Perform standardization processing on the raw one-dimensional voltage data of the common connection point of the power distribution network, converting it into a one-dimensional time-domain feature tensor. The specific operation process is as follows: The collected raw one-dimensional voltage data is standardized and converted into a one-dimensional time-domain feature tensor YF that meets the network input requirements.

[0045] It should be noted that, in the embodiments of the present invention, the standardization process can use Z-Score standardization to adjust the original data into a distribution with a mean of 0 and a standard deviation of 1.

[0046] S102. Construct a dual-branch parallel enhancement architecture, which includes a first branch architecture and a second branch architecture. The first branch architecture outputs a time-frequency domain feature tensor from the one-dimensional time-domain feature tensor through time-frequency weighted singular value decomposition and spatial alignment. The specific operation process is as follows: It should be noted that, addressing the issue that voltage signals collected at the points of common coupling in distribution networks are often contaminated by complex electromagnetic noise, leading to the submergence of weak disturbance characteristics, this module designs a dual-branch parallel enhancement architecture. On one hand, time-frequency weighted singular value decomposition (TFWS) is used to extract the sparse energy fingerprint of the voltage signal in the frequency domain, effectively separating random white noise from disturbance signals with structured characteristics. On the other hand, the Hankel matrix is ​​used to construct the dynamic trajectory of the voltage signal, and the deterministic voltage waveform evolution law is reconstructed through subspace decomposition. Through the physical complementarity of the time-frequency and time domains, high signal-to-noise ratio pure disturbance characteristics are cleaned and restored from the strong noise background, laying the physical foundation for subsequent high-precision classification. The structure of DDSFEM is as follows: Figure 2 As shown.

[0047] First, YF is processed in parallel in the following two branches: The first branch architecture aims to capture the time-frequency energy distribution characteristics under strong noise background. First, YF is input to Time-Frequency Weighted Singular Value Decomposition (TFWS) for processing. Considering that Gaussian white noise or electromagnetic interference often exists in the power grid environment, this step aims to extract the energy fingerprint of the disturbance signal in different frequency bands and obtain the time-frequency singular value tensor Q1, thereby realizing the separation of the energy of the principal component of the disturbance from the energy of the background noise at the physical level. Finally, a spatial alignment (resize) operation is performed on Q1 to unify the spatial resolution, and the time-frequency domain feature tensor Q2 is output. It should be noted that this invention designs a Time-Frequency Weighted Singular Value Decomposition (TFWS) submodule, which is used to accurately capture the time-frequency energy focusing characteristics of weak voltage disturbances under strong background noise interference, and physically separates the effective disturbance signal from random electromagnetic noise through low-rank approximation reconstruction technology. The structure of TFWS is as follows: Figure 3 As shown.

[0048] Specifically, the input one-dimensional time-domain feature tensor YF undergoes a short-time Fourier transform (STFT) to decompose the non-stationary voltage signal into a two-dimensional time-frequency tensor. This step aims to reveal the local variation of signal frequency components over time, thereby unfolding the transient disturbance characteristics that were originally hidden in the one-dimensional voltage waveform onto a two-dimensional plane and establishing an intuitive mapping from time to frequency.

[0049] Will The input is fed to the Energy Focus Unit (EFU), which calculates the local amplitude energy distribution in the time-frequency plane to identify high-energy perturbation regions and low-energy noise backgrounds, thereby obtaining the time-frequency energy weight tensor. .

[0050] In the main path of feature enhancement, and Element-wise multiplication is performed to weight and enhance weak perturbation features and suppress background noise, outputting a weighted time-frequency tensor. .

[0051] Will Perform SVD (Singular Value Decomposition) based on the low-rank approximation principle. The top k largest singular values ​​are retained for low-rank reconstruction, and the random noise corresponding to the small singular values ​​is filtered out to output the time-frequency singular value tensor Q1.

[0052] In the above process, the process of applying TFWS to process YF to obtain Q1 is shown below: in, It is a short-time Fourier transform. For energy-focused calculations, For low-rank reconstruction, by retaining only the first k largest singular values ​​after SVD decomposition and reconstructing the matrix, the best low-rank approximation of the original matrix can be obtained, thereby effectively filtering out random background noise represented by tiny singular values; This is a singular value decomposition.

[0053] It should also be noted that this invention designs an Energy Focus Unit (EFU) submodule to construct a time-frequency attention mechanism based on energy saliency. By calculating local amplitude energy and combining it with Gaussian smoothing, it automatically locks high-energy perturbation regions in the voltage signal and generates a spatial weight mask to suppress background noise, thereby achieving precise localization and enhancement of key perturbation features. The EFU structure is as follows: Figure 4 As shown.

[0054] Specifically, for Performing an element-wise product with itself can widen the gap between the signal and noise, outputting a coarse-grained energy tensor. ; and then using a The Gaussian smoothing kernel will A sliding convolution operation is performed to aggregate local spatial texture information, outputting an energy focusing tensor E1; subsequently, the energy distribution is mapped to normalized weights using a sigmoid activation function, outputting a time-frequency energy weight tensor. This aims to allocate greater attention to the time-frequency regions where disturbances occur.

[0055] In the above process, EFU is applied to Processing yields The process is as follows: in, This is an element-wise product, that is, tensors are multiplied element by element; It is a Gaussian smoothing kernel, a linear smoothing operator based on the normal distribution function, used for convolution operations.

[0056] S103. The second branch architecture constructs a Hankel trajectory tensor from the one-dimensional time-domain feature tensor through a multi-scale sliding window, and performs singular value decomposition to obtain the time-domain tensor; the time-frequency domain feature tensor and the time-domain tensor are concatenated along the channel dimension to obtain a time-frequency dual-domain feature tensor. The specific operation process is as follows: In the second branch architecture, YF constructs the Hankel trajectory tensor D1 via a multi-scale sliding window module. By setting the sliding window length and sliding step size, it achieves a one-dimensional to two-dimensional mapping. This step is essentially a phase space reconstruction of the voltage signal, aiming to reveal the evolution trajectory of the nonlinear dynamics of the power system through high-dimensional embedding. Simultaneously, D1 is processed by SVD to obtain the time-domain singular spectrum tensor D2. A resize operation is performed on D2 to match the dimension, and the time-domain tensor D3 is output.

[0057] By concatenating Q2 from branch one and D3 from branch two along the channel dimension, TF is obtained. This operation achieves physical complementarity between spectral energy features and temporal dynamic features, enabling the model to identify which frequency components it contains and to perceive how the waveform evolves over time.

[0058] In summary, during the above process, the dual-domain subspace feature enhancement module is applied to process YF to obtain TF, as shown in the following equation: .

[0059] in, For spatial alignment operations of bilinear interpolation, This is for splicing operations.

[0060] In this embodiment of the invention, step S200 involves fusing a one-dimensional time-domain feature tensor and a time-frequency dual-domain feature tensor through time-frequency dual interactive guidance to obtain a software-hardware interactive fused feature tensor, including the following steps S201-S203: YF and TF are input into the Time-Frequency Dual-Interaction Guided Fusion Module (TFDIGFM), which outputs the software-hardware interactive fusion feature tensor DG. TFDIGFM addresses the challenge of a heterogeneous gap between one-dimensional time-domain feature tensors and dual-channel frequency-domain feature tensors in power quality composite disturbance analysis, making deep waveform and spectrum fusion difficult. It designs a dual attention interaction mechanism (soft and hard). The core of this mechanism is to capture the global correlation between key triggering features of the time-domain voltage waveform and the frequency-domain harmonic energy distribution using a soft attention mechanism. Simultaneously, a hard attention mechanism is used to directly filter effective spectral channels containing fault information based on the time-domain waveform deviation. This achieves precise alignment and complementarity of heterogeneous features in the physical evolution of disturbances, effectively solving the problem that single-domain features cannot fully characterize composite disturbances. It significantly improves the robustness of features to dynamic grid interference. The time-frequency dual interaction-guided fusion module, for example... Figure 5 As shown.

[0061] S201. Map the one-dimensional time-domain feature tensor to a query tensor, and map the time-frequency dual-domain feature tensor to a key tensor and a value tensor; calculate the tensor product of the query tensor and the key tensor, extract the globally weighted feature vector, and expand it to the same spatial dimension as the time-frequency dual-domain feature tensor to obtain the time-domain and frequency-domain feature fusion soft tensor. The specific operation process is as follows: exist Figure 5 In the soft attention fusion channel, YF is mapped to query tensor Q via linear transformation, aiming to extract time-triggered features at the time of perturbation such as voltage dips, spikes or interruptions; TF is input into a 1×1 convolutional layer and reshaped to map to key tensor K; TF is input into another set of convolutional layers and reshaped to obtain value tensor V.

[0062] Calculate the tensor product of Q and K and perform Softmax processing to generate a soft attention weight tensor A1 that reflects the causal mapping relationship between voltage time-domain waveform abrupt changes and specific frequency band energy; Tensor multiplication is performed between A1 and V to extract weighted features, resulting in a global weighted feature vector U. This vector is then expanded to the same spatial dimension as TF via a broadcast operation, outputting a time-domain and frequency-domain feature fusion soft tensor T1 that incorporates global time-domain context guidance. This allows the model to automatically focus on the corresponding harmonic or interharmonic feature regions in the frequency domain based on voltage abrupt changes in the time domain.

[0063] The process of obtaining T1 by processing YF and TF in the soft attention fusion channel is shown below: Wherein, FC represents a fully connected layer; Broadcasting is a dimension alignment and expansion mechanism widely used in tensor operations in deep learning; Tensor transformation operations, widely used in deep learning models, are used for dimensional reshaping. To normalize the exponential function, the arbitrary real-domain logistic regression vector output by the fully connected layer is mapped to the normalized exponential function through an exponential transformation. Find the range of positive real numbers and force the sum of all elements to be 1. For convolution operations used to generate keys in the attention mechanism; The convolution operation generates values ​​used to generate the value tensors in the attention mechanism. This is a matrix transpose operation.

[0064] S202. The one-dimensional time-domain feature tensor is processed through a fully connected layer, dimension reshaping, and the Sigmoid function to obtain a hard attention weight tensor. The hard attention weight tensor is then multiplied element-wise with the time-frequency dual-domain feature tensor to directly filter the effective channels in the frequency domain, thus obtaining a time-frequency domain feature fusion hard tensor. The specific operation process is as follows: exist Figure 5 In the hard attention fusion channel, YF is input into the fully connected layer and reshape it. Then, it is processed by the Sigmoid activation function to obtain the hard attention weight tensor A2, which aims to quantify the contribution of each frequency domain channel to the power quality disturbance event at the current moment.

[0065] By performing element-wise product operation between A2 and TF, effective channels in the frequency domain are directly filtered out and redundant noise is suppressed using time-domain saliency. The output time-domain and frequency-domain feature fusion hard tensor T2 is then used to achieve hard blocking of background spectral noise that is irrelevant to the current perturbation type, ensuring that the model focuses only on spectral components with physical meaning.

[0066] The process of obtaining T2 by processing YF and TF in the hard attention fusion channel is shown below: in, FC stands for element-wise product, and FC stands for fully connected layer.

[0067] S203. Concatenate the soft tensor of time-domain and frequency-domain feature fusion with the hard tensor of time-domain and frequency-domain feature fusion to obtain the soft-hard interactive fusion feature tensor. The specific operation process is as follows: The features of T1 and T2 are concatenated along the channel dimension to achieve complementary interaction between hard and soft features. The concatenated features are then processed by a single... The convolutional layer performs channel mapping and semantic alignment to obtain the final soft-hard interaction fusion feature tensor DG; The process of concatenating the channel-dimensional features of T1 and T2 to obtain DG is shown below: in, for Convolutional layer.

[0068] In an embodiment of the present invention, step S300 involves multi-granularity semantic mining and aggregation of the software-hardware interaction fusion feature tensor to generate a high-dimensional robust feature tensor with multi-granularity information, including the following steps S301-S304: The DG is input into the Multi-Granularity Semantic Mining and Aggregation Module (MGSMA), which outputs a high-dimensional robust feature tensor (MIHRF) of multi-granularity information. The architecture of the MGSMA module is as follows: Figure 6 As shown S301. The soft-hard interaction fusion feature tensor is input into the global average pooling layer for processing to obtain the global average pooling tensor. A flattening operation is then performed to obtain the global flattened feature tensor. The specific operation process is as follows: exist Figure 6 In this process, the DG is input into the global average pooling layer for processing, and the information is compressed along the spatial dimension to obtain the global average pooling tensor GAPT. Performing a flatten operation on GAPT expands the multidimensional tensor into a one-dimensional tensor, outputting a globally flattened feature tensor FGAPT that contains global macroscopic semantics. This allows the model to ignore random disturbances at the initial moment of voltage perturbation, as shown below: in, This is a tensor dimension reshaping operation used to expand a multidimensional feature tensor into a one-dimensional continuous tensor; Global average pooling, used in this invention, is employed to filter out spatial location interference, retaining only the overall response strength of each channel to a specific type of disturbance. S302. Input the soft-hard interaction fusion feature tensor into the adaptive pooling layer for processing to obtain the adaptive pooling tensor and perform a flattening operation to obtain the adaptive flattened feature tensor. The specific operation process is as follows: exist Figure 6 In branch two, in order to preserve the key local structural information in the feature tensor, the DG input is processed by the Adaptive Pooling layer to obtain the Adaptive Pooling Tensor AAPT, which aims to preserve the key topological shape of the voltage transient waveform in the time-frequency domain. Similarly, a flattening operation is performed on AAPT, outputting an adaptive flattened feature tensor FAAPT that includes local key structures. This prevents the loss of fine time-frequency texture details needed to distinguish complex composite perturbations due to over-compression, as shown below: in, For adaptive pooling, in this invention, it is used to preserve the key local topology of the perturbation signal on the time-frequency tensor while compressing the data.

[0069] S303. The soft-hard interaction fusion feature tensor is processed using the gradient-guided detail refinement method to obtain the gradient-guided refined feature tensor, and then flattened to obtain the detail flattened feature tensor. The specific operation process is as follows: exist Figure 6 In the process, the DG is input into GGDR to obtain GGF. Then, the GGF is flattened to directly expand the pixel value in the two-dimensional feature tensor, and the detailed flattened feature tensor FDG is output, as shown below: in, Specifically, this invention designs a Gradient-Guided Detail Refinement Module (GGDR) to extract subtle features that are easily overlooked in voltage signals. It utilizes texture feature extraction paths to capture high-frequency oscillations and glitches, employs spatial gradient attention paths to locate the start and end edges of voltage abrupt changes, and re-adds these subtle details to the features through residual connections. This prevents the loss of critical fault waveform details in deep computation. GGDR, for example... Figure 7 As shown; The input feature tensor DG is fed into this module and processed in parallel along the following two paths to extract texture details and spatial gradient information, respectively. In this invention: The first path is the texture feature extraction path. To achieve high-frequency feature capture, the DG input is... The texture feature tensor G1 is obtained by processing the data through a depthwise convolutional layer (DepthwiseConv); then G1 is input into... The convolutional layer process yields a fine texture feature tensor G2. Subsequently, G2 undergoes a batch normalization (BN) layer and a ReLU activation function for nonlinear transformation, outputting a refined texture feature tensor G3. This is intended to extract the weak high-frequency oscillation texture superimposed on the power frequency voltage waveform and prevent these micro-features from being smoothly lost during deep feature transmission.

[0070] The second path, the spatial gradient attention path, involves locating salient edge regions in the feature tensor. First, the DG is input into both a channel-dimensional max-pooling layer (Maxpool) and a channel-dimensional average-pooling layer (Avgpool), yielding max-pooled feature tensor G4 and average-pooled feature tensor G5, respectively. Then, G4 and G5 are concatenated to obtain a combined feature tensor G6. Finally, G6 is input... The convolutional layer performs gradient perception with a large receptive field to obtain the gradient feature tensor G7. Finally, G7 is processed by the sigmoid activation function to generate the spatial gradient attention weight tensor G8. The aim is to accurately locate the time-frequency edge of power quality disturbance events through gradient saliency, thereby enhancing the model's ability to perceive the duration of disturbances and waveform abrupt boundaries.

[0071] Finally, G3 and G8 are concatenated to obtain the weighted texture feature tensor G9; G9 is then processed... Convolution yields G9', which is then element-wise added to DG via a residual connection to output the final gradient-guided refined feature tensor GGF, thereby achieving targeted enhancement and complete reconstruction of key transient features in composite perturbation signals.

[0072] The process of applying GGDR to process DG to obtain GGF is shown in the following formula: in, and for Convolutional layers and Convolutional layer Max pooling is a non-linear downsampling operation applied in deep neural networks.

[0073] S304. The obtained global flattened feature tensor, adaptive flattened feature tensor, and detail flattened feature tensor are concatenated along the feature dimension. Through multi-granularity aggregation, a high-dimensional robust feature tensor with multi-granularity information is obtained. The specific operation process is as follows: The FGAPT, FAAPT, and FDG obtained from the above three branches are concatenated along the feature dimension to achieve multi-granularity aggregation of macroscopic, mesoscopic, and microscopic semantics, resulting in the multi-granularity information high-dimensional robust feature tensor MIHRF, as shown below: In an embodiment of the present invention, in step S400, a power quality composite disturbance type identification result is generated by probabilistic mapping of a high-dimensional robust feature tensor based on multi-granularity information through a fully connected classification module, including the following steps S401-S402: The MIHRF is input into the fully connected classification module for probability mapping, generating the power quality composite disturbance type identification result Result. The fully connected classification module is responsible for mapping the MIHRF to the power quality disturbance category space. Through linear weighting of the fully connected layer and probability normalization by Softmax, the matching confidence of the feature tensor with various voltage disturbance modes is quantified, thereby realizing the identification of complex composite disturbance types. The fully connected classification architecture is as follows: Figure 8 As shown.

[0074] S401. Map the multi-granularity information high-dimensional robust feature tensor to the power quality disturbance category space, and perform linear weighting through a fully connected layer to obtain the logistic regression tensor. The specific operation process is as follows: The MIHRF is input into a fully connected layer (FC Layer) for processing to obtain the logistic regression tensor Logits, where Logits contains the original confidence scores for each perturbation category.

[0075] Where W is the trainable parameter matrix and b is the bias value.

[0076] S402. The logistic regression tensor is subjected to probability normalization using an activation function to obtain a probability distribution tensor. The specific operation process is as follows: The Logits are input into the Softmax activation function for probability normalization, resulting in the probability distribution tensor P: .

[0077] S403. Perform a maximum value indexing operation on the probability distribution tensor, select the category index corresponding to the element with the largest value in the probability distribution as the final determination basis, and generate the final power quality composite disturbance type identification result. The specific operation process is as follows: Perform a maximum value indexing operation (Argmax) on P, selecting the category index corresponding to the element with the largest value in the probability distribution as the final judgment basis for the current input signal by the model, and generating the final power quality composite disturbance type identification result Result. This Result corresponds to one of the following 10 preset power quality signal categories: ; Argmax is the maximum value index, which finds the index position corresponding to the maximum value in the tensor.

[0078] Example 3 is an embodiment of the present invention, which provides a method for identifying composite power quality disturbances. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0079] In S100, the original one-dimensional voltage signal from the distribution network point of common coupling (PCC) is acquired and truncated to a length of 1024 sampling points, which is then used as the system input.

[0080] First, the data undergoes Z-score normalization using S101 to adjust it to a distribution with a mean of 0 and a standard deviation of 1, resulting in a one-dimensional time-domain feature tensor YF. The size of YF is... .

[0081] Subsequently, YF is input into the Dual-Domain Subspace Feature Enhancement Module (DDSFEM) for deep feature cleaning and enhancement, resulting in the time-frequency dual-domain feature tensor TF, with a size of [missing value]. Pixels, 2 channels.

[0082] In the DDSFEM module, YF is processed in parallel. In branch one, YF is first input to the Time-Frequency Weighted Singular Value Decomposition (TFWS) submodule for processing, outputting a time-frequency singular value tensor Q1, the size of which is... Pixels, 1 channel; then, a spatial alignment (resize) operation is performed on Q1 to ensure uniform spatial resolution, outputting a time-frequency domain feature tensor Q2, with a size of Pixels, 1 channel. In branch two, YF first constructs the Hankel trajectory tensor D1, setting the embedding dimension to 32 and the sliding step size to 32, resulting in a tensor of size [missing information]. The first pixel, 1 channel D1; then D1 is decomposed using SVD. After SVD decomposition, the first pixel is selected. Low-rank reconstruction is performed on the maximal singular values ​​to obtain the time-domain singular spectrum tensor D2, with a size of . Pixels, 1 channel; finally, to align with the dimensions of branch one, a resize operation is performed on D2, outputting a temporal tensor D3 with a size of... Pixels, 1 channel. Finally, Q2 from branch one and D3 from branch two are concatenated along the channel dimension to achieve physical complementarity between frequency domain and time domain features, thus obtaining the time-frequency dual-domain feature tensor TF, with a size of [missing information]. Pixels, 2 channels.

[0083] In the TFWS submodule, firstly, the YF is converted into a two-dimensional time-frequency tensor using the Short Time Fourier Transform (STFT). Size is Pixels, 1 channel. This tensor is then fed into the Energy Focusing Unit (EFU), which outputs a time-frequency energy weight tensor. Size is Pixels, 1 channel. Next, the two-dimensional time-frequency tensor... with weight tensor Perform element-wise multiplication to achieve weighted enhancement, and output the weighted time-frequency tensor. Size is Pixels, 1 channel. Then, Perform SVD decomposition. After decomposition, select the first... The largest singular values ​​are reconstructed in low rank. Random noise corresponding to the smaller singular values ​​is filtered out and reconstructed. The final output is Q1, which has a size of . Pixels, 1 channel.

[0084] In the EFU submodule, for Energy significance calculation is performed. First, element-wise multiplication is performed on the input tensor to widen the dynamic difference between the signal and noise, outputting a coarse-grained energy tensor. Size is Pixels, 1 channel; then, using a Gaussian smoothed kernel Perform sliding convolution to aggregate local spatial texture information and fill energy holes, outputting an energy focusing tensor E1 of size [value missing]. Pixels, 1 channel; finally, E1 is mapped via the Sigmoid function to output a normalized time-frequency energy weight tensor. Its size is Pixels, 1 channel.

[0085] In S200, the size is YF and size are A pixel-level, 2-channel TF input is fed into a time-frequency dual interactive guided fusion module (TFDIGFM), which outputs a hardware-software interactive fusion feature tensor DG, with a size of [missing information]. Pixels, 32 channels.

[0086] In the TFDIGFM module, the input YF and TF are processed as follows: In the soft attention fusion channel, firstly, the size is... The YF input fully connected layer is linearly mapped to obtain the query tensor Q, which has a size of . At the same time, the size is One pixel, 2-channel TF input The convolutional layer expands the channels to 128 and performs a reshape operation to obtain the bond tensor K, which has a size of . Then, the TF is input into another set of convolutional layers while keeping the number of channels constant, and a reshape operation is performed to obtain a value tensor V with a size of . Subsequently, the tensor product of Q and the transpose of K is calculated and Softmax normalized to obtain the soft attention weight tensor A1, whose size is... Next, tensor multiplication is performed between A1 and V to obtain a globally weighted feature vector U. This U is then expanded to the original spatial resolution using a broadcast operation, resulting in a time-domain and frequency-domain feature fusion soft tensor T1, with a size of [missing value]. Pixels, 2 channels.

[0087] In the hard attention fusion channel, firstly, the size is... The YF input fully connected layer performs feature mapping and dimensionality expansion, mapping the data dimension to 8192 dimensions; subsequently, a reshape operation and sigmoid activation are performed to reconstruct the one-dimensional vector into a hard attention weight tensor A2, with a size of . Pixels, 2 channels; then, A2 is multiplied element-wise by TF to obtain the time-domain and frequency-domain feature fusion hard tensor T2, whose size is... Pixels, 2 channels.

[0088] Finally, the Concat operation is used to concatenate T1 and T2 along the channel dimension to obtain an intermediate feature map with a size of [size missing]. Pixels, 4 channels; then it is converted into one The convolutional layer performs channel mapping and feature integration, ultimately yielding a hardware-software interactive fusion feature tensor, DG, with a size of [missing value]. Pixels, 32 channels.

[0089] In S300, a 64×64 pixel, 32-channel Data Gathering (DG) is input to the Multi-Granularity Semantic Mining and Aggregation Module (MGSMA). Within MGSMA, the DG is fed into three parallel branches for multi-granularity feature extraction, and finally aggregated to output a high-dimensional robust feature tensor MIHRF, which is a one-dimensional tensor with a length of 131616.

[0090] In MGSMA, the input DG is processed in parallel as follows: ① Input the 64×64 pixel, 32-channel DG into the Global Average Pooling layer and compress and aggregate the features along the spatial dimension to obtain the Global Average Pooling Tensor GAPT, which is 1×1 pixel and 32 channels in size; then perform a flattening operation on GAPT to expand it into a one-dimensional tensor to obtain the Global Flattened Feature Tensor FGAPT containing global macro-semantics, which has a length of 32.

[0091] ② Input the DG into the Adaptive Pooling layer, and set the output size to 4×4 pixels to preserve key local structures, resulting in the Adaptive Pooling Tensor AAPT, which is 4×4 pixels and has 32 channels. Then, flatten the AAPT to obtain the Adaptive Flattened Feature Tensor FAAPT, which contains key local structures and has a length of [missing information]. .

[0092] ③ First, input DG into GGDR to obtain a gradient-guided refined feature tensor GGF with a size of 64×64 pixels and 32 channels; then perform a flattening operation on GGF to directly unfold and arrange the values ​​of each pixel, outputting a detail flattened feature tensor FDG containing complete pixel-level information, with a length of .

[0093] Finally, the FGAPT, FAAPT and FDG obtained from the above three branches are concatenated along the feature dimension to achieve multi-granularity aggregation of macro, meso and micro semantics, thus obtaining the final multi-granularity information high-dimensional robust feature tensor MIHRF, with a total length of 32+512+131072=131616.

[0094] In the GGDR submodule, the DG is first input into the module, and then the DG is processed in parallel along two paths: In the texture feature extraction path, DG is first input to... The deep convolutional layer process yields a texture feature tensor G1, with a size of [missing value]. Pixels, 32 channels; then, G1 enters... The convolutional layer performs inter-channel information exchange to obtain a fine texture feature tensor G2, which is still of size 1. Pixels, 32 channels; then G2 undergoes batch normalization and ReLU activation in sequence, outputting the texture refinement feature tensor G3.

[0095] In the spatial gradient attention path, DG passes through a channel-dimension max pooling layer and a channel-dimension average pooling layer, respectively, outputting a max pooling feature tensor G4 and an average pooling feature tensor G5, both of which are of similar size. Pixels, 1 channel. Next, G4 and G5 are concatenated to obtain the combined feature tensor G6, with a size of [missing information]. Pixels, 2 channels; G6 then enters Large kernel convolutional layer, output gradient feature tensor G7, with size . Pixels, 1 channel; then processed by the Sigmoid activation function to generate a spatial gradient attention weight tensor G8, with a size of [missing value]. Pixels, 1 channel.

[0096] Finally, G3 and G8 are concatenated along the channel dimension to obtain the weighted texture feature tensor G9, which has a size of [size missing]. Pixels, 33 channels. To achieve residual connection, the G9 is processed through... The convolution yields G9', whose dimension is... Pixels, 32 channels; then add G9' and DG element-wise to output the gradient-guided refined feature tensor GGF, which has a size of Pixels, 32 channels.

[0097] In S400, a MIHRF of length 131616 is input to the fully connected classification module. Within the module, the MIHRF undergoes probability mapping and activation processing via a fully connected layer, ultimately generating the power quality composite disturbance type identification result Result.

[0098] In the fully connected classification module, the input MIHRF is processed as follows: Execute S401: Input a MIHRF of length 131616 into a fully connected layer (FCLayer), using a size of The trainable weight matrix W and the bias value b of length 10 are linearly transformed. After calculation, the high-dimensional feature space is mapped to the class space, resulting in a logistic regression tensor Logits containing the original confidence scores of each perturbation class, with a length of 10.

[0099] Execute S402: Input Logits into the Softmax activation function for probability normalization, map the logistic regression values ​​to positive real numbers in the interval [0,1], and the sum of all elements is 1, thereby obtaining the probability distribution tensor P representing the predicted probability of each power quality disturbance category, with a length of 10.

[0100] Execute S403: Perform the Argmax operation on P, selecting the element with the largest value from 10 probability values. This patent presupposes a total of 10 categories of power quality signals to be identified, specifically including: ① Normal signal; ② Voltage sag, ③ Voltage swell, ④ Voltage interruption, ⑤ Harmonic, ⑥ Voltage flicker, ⑦ Oscillating transient; ⑧ Voltage sag + harmonic, ⑨ Voltage swell + harmonic, ⑩ Oscillating transient + harmonic. The corresponding category index is used as the final judgment basis for the model on the current input signal, generating the final power quality composite disturbance type identification result Result.

[0101] Assume the input signal is a composite disturbance sample consisting of a voltage sag and harmonics. After Softmax normalization, the generated value of P with a length of 10 might be: In this probability distribution, the probability of voltage sag ② corresponding to index 2 is 0.15, and the probability of harmonic ⑤ corresponding to index 5 is 0.08, which reflects that the model has identified a single disturbance component; the probability of voltage sag + harmonic ⑧ corresponding to index 8 is 0.74, which is the maximum value in the tensor.

[0102] Subsequently, the Argmax operation selects the index position where the maximum value of 0.74 is located, which is the eighth type of voltage sag + harmonic. The final output Result is index 8, thus completing the correct identification of the composite disturbance signal.

[0103] Example 4 is an embodiment of the present invention, and the above is an illustrative scheme of a power quality complex disturbance identification method. It should be noted that the technical solution of a power quality complex disturbance identification system and the technical solution of the power quality complex disturbance identification method described above belong to the same concept. Details not described in detail in the technical solution of the power quality complex disturbance identification system in this embodiment can be found in the description of the technical solution of the power quality complex disturbance identification method described above.

[0104] This embodiment provides a power quality composite disturbance identification system, including: a dual-domain subspace feature enhancement module, a time-frequency dual interactive guided fusion module, a multi-granularity semantic mining and aggregation module, and a fully connected classification module; The dual-domain subspace feature enhancement module standardizes the collected signals from the common connection point of the distribution network to obtain a one-dimensional time-domain feature tensor, and then outputs the one-dimensional time-domain feature tensor as a time-frequency dual-domain feature tensor through dual-domain subspace feature enhancement. The time-frequency dual-interaction guided fusion module performs time-frequency dual-interaction guided fusion on the one-dimensional time-domain feature tensor and the time-frequency dual-domain feature tensor to obtain the software-hardware interactive fusion feature tensor. The multi-granularity semantic mining and aggregation module performs multi-granularity semantic mining and aggregation on the soft-hard interaction fusion feature tensor to generate a multi-granularity information high-dimensional robust feature tensor. The fully connected classification module generates power quality composite disturbance type identification results by performing probability mapping through fully connected classification based on multi-granularity information high-dimensional robust feature tensors.

[0105] This embodiment also provides an electronic device applicable to a power quality composite disturbance identification method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the power quality composite disturbance identification method proposed in the above embodiment.

[0106] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements a power quality composite disturbance identification method as proposed in the above embodiments.

[0107] The storage medium proposed in this embodiment and the method for identifying composite power quality disturbances proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0108] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0109] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for identifying composite power quality disturbances, characterized in that: include, After standardizing the acquired signals from the common connection point of the distribution network, a one-dimensional time-domain feature tensor is obtained. The one-dimensional time-domain feature tensor is then output as a time-frequency dual-domain feature tensor through dual-domain subspace feature enhancement. The one-dimensional time-domain feature tensor and the time-frequency dual-domain feature tensor are fused by time-frequency dual interaction guidance to obtain the software-hardware interactive fusion feature tensor. Multi-granularity semantic mining and aggregation are performed on the aforementioned hardware-software interaction fusion feature tensor to generate a multi-granularity information high-dimensional robust feature tensor; Based on the high-dimensional robust feature tensor with multi-granularity information, probability mapping is performed through fully connected classification to generate power quality composite disturbance type identification results.

2. The method for identifying composite power quality disturbances as described in claim 1, characterized in that, The enhancement of the dual-domain subspace features includes: The raw one-dimensional voltage data of the common connection point of the power distribution network is standardized and converted into a one-dimensional time-domain feature tensor. A dual-branch parallel enhancement architecture is constructed, which includes a first branch architecture and a second branch architecture; The first branch architecture outputs a time-frequency domain feature tensor from the one-dimensional time-domain feature tensor through time-frequency weighted singular value decomposition and spatial alignment; The second branch architecture constructs the Hankel trajectory tensor from the one-dimensional time-domain feature tensor through a multi-scale sliding window, and obtains the time-domain tensor by performing singular value decomposition. The time-frequency domain feature tensor and the time domain tensor are concatenated along the channel dimension to obtain the time-frequency dual-domain feature tensor.

3. The method for identifying composite power quality disturbances as described in claim 2, characterized in that, The time-frequency weighted singular value decomposition includes: Short-time Fourier transform is performed on the one-dimensional time-domain feature tensor to decompose the non-stationary voltage signal into a two-dimensional time-frequency tensor. The local amplitude energy distribution in the time-frequency plane is calculated from the two-dimensional time-frequency tensor, the perturbation region and noise background are identified, and the time-frequency energy weight tensor is obtained. The weighted time-frequency tensor is obtained by multiplying the two-dimensional time-frequency tensor element by element with the time-frequency energy weight tensor; The time-frequency singular value tensor is decomposed using singular value decomposition, and random noise in the weighted time-frequency tensor is filtered out to output the time-frequency singular value tensor.

4. The method for identifying composite power quality disturbances as described in claim 3, characterized in that, The time-frequency dual interactive guidance fusion includes: Map the one-dimensional time-domain feature tensor to the query tensor, and the time-frequency dual-domain feature tensor to the key tensor and the value tensor; Calculate the tensor product of the query tensor and the key tensor, extract the globally weighted feature vector, and expand it to the same spatial dimension as the time-frequency dual-domain feature tensor to obtain the time-domain and frequency-domain feature fusion soft tensor; The one-dimensional time-domain feature tensor is processed by a fully connected layer, dimension reshaping and sigmoid function to obtain a hard attention weight tensor. The hard attention weight tensor is then multiplied element-wise with the time-frequency dual-domain feature tensor to directly filter the effective channels in the frequency domain and obtain the time-frequency domain feature fusion hard tensor. By concatenating the soft tensor of time-domain and frequency-domain feature fusion with the hard tensor of time-domain and frequency-domain feature fusion, a soft-hard interactive fusion feature tensor is obtained.

5. The method for identifying composite power quality disturbances as described in claim 4, characterized in that, The multi-granularity semantic mining and aggregation includes: The soft-hard interaction fusion feature tensor is input into the global average pooling layer for processing to obtain the global average pooling tensor and then flattened to obtain the global flattened feature tensor. The soft-hard interaction fusion feature tensor is input into the adaptive pooling layer for processing to obtain the adaptive pooling tensor and then a flattening operation is performed to obtain the adaptive flattened feature tensor. The soft-hard interaction fusion feature tensor is used to obtain a gradient-guided refinement feature tensor by using the gradient-guided detail refinement method, and then a flattening operation is performed to obtain a detail flattening feature tensor. The obtained global flattened feature tensor, adaptive flattened feature tensor, and detailed flattened feature tensor are concatenated along the feature dimension, and a high-dimensional robust feature tensor with multi-granularity information is obtained through multi-granularity aggregation.

6. The method for identifying composite power quality disturbances as described in claim 5, characterized in that, The gradient-guided detail refinement method includes: The gradient-guided detail refinement method processes the first and second paths in parallel. In the first path, the soft and hard interaction fusion feature tensor is processed through a deep convolutional layer, and the resulting texture feature tensor is nonlinearly transformed to generate a texture refinement feature tensor. In the second path, the soft and hard interaction fusion feature tensor is processed by the max pooling layer and the average pooling layer respectively and then concatenated to obtain the combined feature tensor. The combined feature tensor is then subjected to gradient sensing to generate the spatial gradient attention weight tensor. The texture refinement feature tensor is concatenated with the spatial gradient attention weight tensor to obtain a weighted texture feature tensor. The weighted texture feature tensor is then convolved and added element-wise with the soft-hard interaction fusion feature tensor to obtain the gradient-guided refinement feature tensor.

7. The method for identifying composite power quality disturbances as described in claim 6, characterized in that, The probability mapping via the fully connected classification module includes: The multi-granularity information high-dimensional robust feature tensor is mapped to the power quality disturbance category space, and linearly weighted through a fully connected layer to obtain the logistic regression tensor; The logistic regression tensor is subjected to probability normalization using an activation function to obtain a probability distribution tensor; Perform a maximum value indexing operation on the probability distribution tensor, select the category index corresponding to the element with the largest value in the probability distribution as the final judgment basis, and generate the final power quality composite disturbance type identification result.

8. A power quality composite disturbance identification system, employing the power quality composite disturbance identification method as described in any one of claims 1 to 7, characterized in that, include: Dual-domain subspace feature enhancement module, time-frequency dual-interaction guided fusion module, multi-granularity semantic mining and aggregation module, and fully connected classification module; The dual-domain subspace feature enhancement module standardizes the collected signals from the common connection point of the distribution network to obtain a one-dimensional time-domain feature tensor, and then outputs the one-dimensional time-domain feature tensor as a time-frequency dual-domain feature tensor through dual-domain subspace feature enhancement. The time-frequency dual-interaction guided fusion module performs time-frequency dual-interaction guided fusion on the one-dimensional time-domain feature tensor and the time-frequency dual-domain feature tensor to obtain the software-hardware interactive fusion feature tensor. The multi-granularity semantic mining and aggregation module performs multi-granularity semantic mining and aggregation on the soft-hard interaction fusion feature tensor to generate a multi-granularity information high-dimensional robust feature tensor. The fully connected classification module generates power quality composite disturbance type identification results by performing probability mapping through fully connected classification based on multi-granularity information high-dimensional robust feature tensors.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the power quality composite disturbance identification method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the power quality composite disturbance identification method according to any one of claims 1 to 7.

Images