Wavelet transform and lightweight deep learning-based power distribution network area voltage transient event detection method and system
By combining wavelet transform and lightweight deep learning, the real-time and accuracy issues of voltage transient over-limit detection are solved, achieving fast and accurate voltage transient event detection, especially with high robustness in complex disturbance environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- STATE GRID HUNAN ELECTRIC POWER COMPANY LIMITED
- Filing Date
- 2026-01-19
- Publication Date
- 2026-05-12
AI Technical Summary
Existing voltage transient over-limit detection methods are insufficient in terms of response speed and accuracy, especially under complex disturbances, making it difficult to achieve fast and accurate detection.
Wavelet transform is used to extract energy entropy feature vectors, and a lightweight deep learning classification network is combined to determine the type of voltage transient events. Dynamic time warping algorithm is used to calibrate the start and end times of events, and a detection system based on wavelet transform and lightweight deep learning is constructed.
It achieves rapid and accurate detection of voltage transient events, with a single detection process taking a stable time of less than 8 milliseconds and a positioning error of no more than 0.1 milliseconds, thus improving the real-time performance and accuracy of detection.
Smart Images
Figure CN122020369A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power quality analysis and protection control technology in power systems, and in particular to a method and system for detecting voltage transient events in distribution network areas based on wavelet transform and lightweight deep learning. Background Technology
[0002] With the large-scale integration of new energy power generation, distributed power sources, and nonlinear loads into the power system, transient over-limit events such as voltage dips and rises occur frequently, placing higher demands on fast and accurate detection methods.
[0003] Traditional voltage over-limit detection methods are mostly based on calculating the effective voltage value or using a second-order generalized integrator for fundamental frequency extraction and phase synchronization. However, the effective value algorithm suffers from computational delay and is difficult to capture instantaneous changes; integrator methods such as SOGI / DSOGI exhibit response lag and estimation bias under voltage changes and harmonic interference.
[0004] Existing improved solutions, such as the published patent CN120879642A, use wavelet transform to locate the moment of voltage abrupt change. However, its core still relies on the direct comparison of wavelet coefficients with a fixed threshold, lacking intelligent learning capabilities and having weak generalization ability to complex disturbances. Another patent, CN120855539A, proposes a system-level voltage over-limit mitigation scheme, but its mitigation effectiveness is limited by the speed and accuracy of the underlying detection technology.
[0005] Therefore, there is an urgent need for a new method for detecting voltage transient over-limits that can balance real-time performance, accuracy, and robustness. Summary of the Invention
[0006] The purpose of this invention is to provide a method and system for detecting voltage transient events in distribution network areas based on wavelet transform and lightweight deep learning, so as to solve the problems of slow response speed and insufficient accuracy under complex disturbances in existing detection technologies.
[0007] To achieve the above objectives, the present invention adopts the following technical solution:
[0008] A method for detecting voltage transient events in distribution network transformer areas based on wavelet transform and lightweight deep learning includes:
[0009] Step S1: Preprocess and perform multi-scale wavelet transform on the three-phase power grid voltage signal of the distribution network area to extract the wavelet energy entropy feature vector;
[0010] Step S2: Construct and train a lightweight deep learning classification network, using the wavelet energy entropy feature vector as input, and output the probability of various voltage transient events (such as voltage dips and voltage rises) to complete the determination of voltage transient event types.
[0011] Step S3: For the voltage over-limit transient events that are determined to be voltage over-limit, calibrate their start and end times based on the dynamic time warping algorithm;
[0012] Step S4: Output the detection results, including the voltage transient event type and its confidence level, and the start and end times of the voltage over-limit transient event; where the confidence level of the voltage transient event type refers to the probability that the lightweight deep learning classification network determines the type of the current event.
[0013] In one possible implementation, step S1 specifically includes:
[0014] S11: Acquire three-phase power grid voltage signals and perform noise reduction and normalization preprocessing;
[0015] S12: Perform multi-level discrete wavelet decomposition on the preprocessed signal to obtain the high-frequency detail coefficients and low-frequency approximation coefficients of each level;
[0016] S13: Calculate the energy and wavelet energy entropy of each frequency band;
[0017] S14: Combine the energy and energy entropy of each frequency band to form a multi-scale wavelet energy entropy feature vector.
[0018] In one possible implementation, the energy calculation formula for each frequency band in step S13 is as follows:
[0019] ;
[0020] in, For the first Energy of each frequency band Representing the The first frequency band Wavelet coefficients, For the first The coefficient length of each frequency band;
[0021] The formula for calculating wavelet energy entropy is:
[0022] in, For wavelet energy entropy, For the first The energy of each frequency band accounts for a portion of the total energy. The ratio, i.e. ; is the number of frequency bands, where This is equal to the number of levels in the discrete wavelet decomposition. In one possible implementation, step S2 specifically includes:
[0023] In one possible implementation, the lightweight deep learning classification network includes, in sequence: an input layer, a structured feature remodeling layer, a multi-scale deep separable convolutional fusion module, a joint attention module, a global average pooling layer, and an output layer.
[0024] The structured feature renormalization layer is used to rearrange the wavelet energy entropy feature vector according to phase, frequency band and statistics to form a three-dimensional feature tensor, and generate at least one set of interphase differential feature channels in the three-dimensional feature tensor to characterize the three-phase coupling relationship;
[0025] The multi-scale depthwise separable convolutional fusion module includes at least two parallel depthwise separable convolutional branches. Each depthwise separable convolutional branch includes a depthwise separable convolutional layer, a batch normalization layer, and a ReLU activation function connected in sequence. Different depthwise separable convolutional branches use different kernel sizes and / or different dilatation rates to extract transient discriminative features at different scales. The outputs of each depthwise separable convolutional branch are fused through pointwise convolution to obtain a fused feature map. The fused feature map is superimposed on the input features of the multi-scale depthwise separable convolutional fusion module through residual connections.
[0026] The joint attention module includes a channel attention submodule and a frequency band attention submodule. The channel attention submodule generates channel weights based on the channel statistical descriptors obtained by global average pooling, and the frequency band attention submodule generates frequency band weights based on the statistical descriptors of the frequency band dimension. The channel weights and frequency band weights are multiplied by the corresponding dimensions of the input features of the joint attention module to strengthen key transient features and suppress redundant features, thereby obtaining a weighted three-dimensional feature tensor.
[0027] The output layer includes a confidence calibration unit and a Softmax classifier, which output the probability of various voltage transient events.
[0028] In one possible implementation, the training process of the lightweight deep learning classification network employs a wavelet domain-specific loss function, and noisy data is injected into the training dataset.
[0029] In one possible implementation, step S3 specifically includes:
[0030] S31: When the lightweight deep learning classification network determines that the event is a voltage over-limit transient event, extract the high-frequency wavelet coefficient sequence of the event period;
[0031] S32: Perform dynamic time warping calculation on the high-frequency wavelet coefficient sequence and the preset typical transient disturbance template, and output the dynamic time warping cost matrix;
[0032] S33: Based on the dynamic time warping cost matrix, find the optimal path and align the distortion points of the high-frequency wavelet coefficient sequence with the preset typical transient disturbance template through the optimal path;
[0033] S34: Determine the start and end points of voltage over-limit transient events based on aligned distortion points.
[0034] In one possible implementation, the preset typical transient disturbance template in step S32 includes three core templates: voltage sag, voltage swell, and instantaneous interruption. Each template is generated by statistical modeling of the high-frequency wavelet coefficient sequence of no less than 1,000 sets of standard transient event samples. The templates support online updates. When a new type of transient limit-breaking event is detected, its high-frequency wavelet coefficient features are automatically extracted and added to the template library.
[0035] Secondly, this application provides a distribution network transformer area voltage transient event detection system based on wavelet transform and lightweight deep learning, including: a memory and a processor;
[0036] The memory is used to store computer programs;
[0037] The processor is used to invoke the computer program to execute the method described above.
[0038] Thirdly, this application provides a computer-readable storage medium storing a computer program that, when run on an electronic device, causes the electronic device to perform the method described above.
[0039] Fourthly, this application provides a computer program product, including a computer program that, when run on an electronic device, causes the electronic device to perform the method described above.
[0040] The specific implementation methods of the second to fourth aspects of this application can refer to the implementation methods of the first aspect, and will not be elaborated here.
[0041] The present invention has the following beneficial effects:
[0042] This invention achieves rapid and accurate detection of voltage transient events by fusing wavelet transform with lightweight deep learning. The introduction of wavelet energy entropy feature vectors significantly reduces model learning difficulty and data dependence. The lightweight network design ensures that the complete detection process for a single instance remains stable within 8 milliseconds while maintaining accuracy, thus guaranteeing a real-time response at the 8-millisecond level. Combined with the Dynamic Time Warping (DTW) algorithm, the start and end times of voltage transient events can be accurately calibrated, precisely aligning the distortion points of the actual signal and the template, thereby achieving accurate calibration. The positioning error of the start and end times of voltage transient events does not exceed 0.1 milliseconds, achieving a positioning accuracy at the 0.1-millisecond level. This forms a complete technical closed loop from detection to positioning, effectively enhancing the engineering application value of the system. Attached Figure Description
[0043] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:
[0044] Figure 1 This is a schematic diagram of a method flow in one embodiment of this application;
[0045] Figure 2 This is a schematic diagram of the wavelet feature extraction process in one embodiment of this application;
[0046] Figure 3 This is a schematic diagram of a lightweight deep learning classification network structure in one embodiment of this application. Detailed Implementation
[0047] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.
[0048] This embodiment discloses a method for detecting voltage transient events in distribution network transformer areas based on wavelet transform and lightweight deep learning, such as... Figures 1-3 As shown, firstly, wavelet transform is performed on the three-phase voltage signals of the distribution network area to extract energy entropy features. Then, a lightweight deep learning classification network is used for rapid classification. Finally, the start and end points of out-of-limit events are precisely located. Specifically, this includes:
[0049] Step S1: Preprocess and perform multi-scale wavelet transform on the three-phase grid voltage signal of the distribution network area to extract the wavelet energy entropy feature vector.
[0050] Preferably, this step specifically includes:
[0051] S11: Acquire three-phase grid voltage signals (ua, ub, uc) and perform noise reduction and normalization preprocessing.
[0052] S12: Perform multi-level (e.g., 4-level) discrete wavelet decomposition on the preprocessed signal to obtain the high-frequency detail coefficients of each level (e.g., Figure 2 (D1~D4) and low-frequency approximation coefficients (such as...) Figure 2 (A4).
[0053] Among them, the wavelet basis functions can be Db4 and Sym8.
[0054] Furthermore, the number of layers in discrete wavelet decomposition It can adaptively adjust according to the sampling frequency of the power grid voltage signal, with an adjustment range of 3-6 layers; when the sampling frequency is ≥10kHz, The value is set to 5-6 layers; when the sampling frequency is between 5-10kHz, The value is set to 4 layers; when the sampling frequency is < 5kHz, The number of layers is set to 3. This scheme employs an adaptive mechanism of "sampling frequency - number of decomposition layers". The higher the sampling frequency, the richer the high-frequency details contained in the signal, and multiple layers are decomposed to capture features; when the sampling frequency is low, the number of layers is reduced to avoid redundant calculations.
[0055] Among them, the low-frequency approximation coefficient corresponds to the lower-frequency, gently changing components in the original voltage signal, mainly reflecting the "overall trend" of the signal (such as the fundamental component of the grid voltage and the continuous characteristics during stable operation). After layer discrete wavelet decomposition, the low-frequency approximation coefficients retain the low-frequency signal characteristics through "low-pass filtering," and their frequency range can be up to the sampling frequency of the original signal. ( To decompose the hierarchy, from 1 to As the decomposition levels increase, the frequency range corresponding to the low-frequency approximation coefficients gradually decreases, focusing on the core stable trend of the signal.
[0056] High-frequency detail coefficients correspond to the higher-frequency, rapidly changing components in the original voltage signal, primarily reflecting the signal's "abrupt characteristics" (such as the instantaneous changes in voltage sags / sags, disturbances, and noise). These high-frequency detail coefficients are high-frequency signal features separated through "high-pass filtering." Each decomposition layer yields a corresponding high-frequency detail coefficient for a given frequency band, with a frequency range that can be a fraction of the original signal's sampling frequency. The higher the level, the higher the frequency corresponding to the high-frequency detail coefficient, and the better it can capture rapid changes in the signal.
[0057] S13: Calculate the energy value of each frequency band (e.g.) Figure 2 E_D1~E_A4) and wavelet energy entropy.
[0058] The energy calculation formulas for each frequency band are as follows:
[0059] ;
[0060] in, For the first Energy of each frequency band Representing the The first frequency band Wavelet coefficients, For the first The coefficient length of each frequency band; the energy intensity of each frequency band can be quantified through this formula.
[0061] The formula for calculating wavelet energy entropy is:
[0062] in, For wavelet energy entropy, For the first The energy of each frequency band accounts for a portion of the total energy. The ratio, i.e. ; is the number of frequency bands, where It equals the number of levels in the discrete wavelet decomposition. Wavelet energy entropy can reflect the uniformity of signal energy distribution across frequency bands, highlighting the characteristic differences of transient disturbances.
[0063] S14: Combine the energy and energy entropy of each frequency band to form a multi-scale wavelet energy entropy feature vector.
[0064] The above steps can transform the original three-phase power grid voltage signal into a feature vector that can be efficiently recognized by a lightweight deep learning classification network.
[0065] Step S2: Construct and train a lightweight deep learning classification network, using the wavelet energy entropy feature vector as input, and output the probability of various voltage transient events (such as voltage dips and voltage rises) to complete the determination of voltage transient event types.
[0066] This step enables rapid and accurate classification of voltage transient events by constructing a dedicated lightweight network.
[0067] Preferably, in this step, construct as follows Figure 3 The lightweight deep learning classification network shown includes, in sequence: an input layer, a structured feature remodeling layer, a multi-scale deep separable convolutional fusion module, a joint attention module, a global average pooling layer, and an output layer.
[0068] The structured feature reshaping layer is used to perform structured transformation and adaptation processing on the input wavelet energy entropy feature vector: the wavelet energy entropy feature vector is rearranged according to phase, frequency band and statistics to form a three-dimensional feature tensor, and at least one set of interphase difference feature channels is generated in the three-dimensional feature tensor. The interphase difference feature channels can explicitly characterize the three-phase coupling, imbalance and interphase abrupt differences, and improve the ability of the subsequent network to identify the three-phase transient features. For example, the dimension of the rearranged three-dimensional feature tensor is the number of phases × the number of frequency bands × the number of statistics, for example: 3×6×2. If 3 interphase difference channels are added, the final dimension of the three-dimensional feature tensor is 3×6×(2+3)).
[0069] The multi-scale depthwise separable convolutional fusion module includes at least two parallel depthwise separable convolutional branches. Each depthwise separable convolutional branch includes a depthwise separable convolutional layer, a batch normalization layer (used to accelerate training and suppress overfitting), and a ReLU activation function (introducing nonlinearity to enhance feature representation) connected in sequence. Different depthwise separable convolutional branches use different kernel sizes and / or different dilatation rates to extract transient discriminative features at different scales. The outputs of each depthwise separable convolutional branch are fused through pointwise convolution to obtain a fused feature map. The fused feature map is superimposed on the input features of the multi-scale depthwise separable convolutional fusion module through residual connections, which further reduces the parameter scale and improves training stability while ensuring the feature extraction effect.
[0070] The joint attention module includes a channel attention submodule and a frequency band attention submodule. The channel attention submodule generates channel weights based on channel statistical descriptors obtained through global average pooling, enhancing key channel features. The frequency band attention submodule generates frequency band weights based on statistical descriptors of the frequency band dimension, enhancing key frequency band features. The channel weights and frequency band weights are multiplied by the corresponding dimensions of the input features of the joint attention module, i.e., the channel weights are multiplied by the channel dimension of the input tensor (channel-wise weighting), and the frequency band weights are multiplied by the frequency band dimension of the input tensor (frequency band-wise weighting). The two weighting operations are performed sequentially, first channel weighting, then frequency band weighting, or vice versa, without changing the tensor dimension. This yields a weighted three-dimensional feature tensor. This step enhances key transient features and suppresses redundant features, ultimately resulting in a three-dimensional feature tensor that enhances key features and suppresses redundancy, thereby improving the model's robustness to noise, harmonics, and other disturbances.
[0071] The output layer includes a confidence calibration unit and a Softmax classifier, which output the probability of various voltage transient events.
[0072] The confidence calibration unit is deployed after the one-dimensional feature vector output by the global average pooling layer and before the Softmax classifier to address the mismatch between the model's predicted probability and actual classification accuracy. It can include basic calibration: scaling the feature vector using a learnable temperature coefficient to correct the overall bias of the Softmax output. It can also include scene calibration: combining the category characteristics of voltage transient events, optimizing the confidence output for different event types through category weight correction and threshold filtering, ultimately making the predicted confidence more closely resemble the true accuracy, facilitating threshold decision-making and engineering alerts.
[0073] Preferably, the network training employs a wavelet domain-specific loss function and injects noisy data.
[0074] The training process of the lightweight deep learning classification network employs a wavelet domain-specific loss function, and noisy data is injected into the training dataset to enhance robustness. This noisy data may include Gaussian noise, impulse noise, and power grid harmonic interference noise. In some embodiments, the signal-to-noise ratio of the Gaussian noise ranges from 20 to 40 dB, the amplitude of the impulse noise does not exceed 5% of the peak value of the training data, and the power grid harmonic interference noise includes 3rd, 5th, and 7th harmonic components.
[0075] Step S3: For the voltage over-limit transient event determined to be a voltage over-limit event, calibrate its start and end times based on the dynamic time warping algorithm.
[0076] This step enables high-precision location of the start and end points for the categorized out-of-limit events.
[0077] Preferably, this step specifically includes:
[0078] S31: When the lightweight deep learning classification network determines that the event is a voltage over-limit transient event, extract the high-frequency wavelet coefficient sequence of the event period.
[0079] High-frequency wavelet coefficients are more sensitive to signal abrupt changes and are key to locating the start and end times.
[0080] S32: Perform dynamic time warping calculation on the high-frequency wavelet coefficient sequence and the preset typical transient disturbance template, and output the dynamic time warping cost matrix.
[0081] Dynamic Time Warping (DTW) can handle temporal distortions between signals and templates (such as differences in event duration or different mutation rhythms).
[0082] Among them, the preset typical transient disturbance templates include three core templates: voltage sag, voltage swell, and instantaneous interruption. Each type of template is generated by statistical modeling of high-frequency wavelet coefficient sequences of no less than 1,000 sets of standard transient event samples. The templates support online updates. When a new type of transient limit-breaking event is detected, its high-frequency wavelet coefficient features are automatically extracted and added to the template library.
[0083] S33: Based on the dynamic time warping cost matrix, find the optimal path and align the distortion points of the high-frequency wavelet coefficient sequence with the preset typical transient disturbance template through the optimal path.
[0084] S34: Determine the start and end points of voltage over-limit transient events based on aligned distortion points.
[0085] Step S4: Output the detection results, including the voltage transient event type (such as normal voltage, voltage drop, voltage rise, etc.) and its confidence level (obtained by the confidence level calibration unit and Softmax classifier from the probability output by the output layer in step S2), and the start and end times of the voltage limit-breaking transient event (with a positioning accuracy of 0.1 milliseconds); where the confidence level of the voltage transient event type refers to the probability that the lightweight deep learning classification network determines the type of the current event.
[0086] Thus, one detection process is completed. Experiments show that the above scheme takes a stable time of less than 8 milliseconds for a single detection process, meeting the requirements for real-time detection.
[0087] It should be understood that the above numbering S1 to S4 is only used to distinguish and facilitate the expression of different steps, and does not necessarily constitute a restriction on the execution order between the steps.
[0088] The technical solution provided in this application integrates multi-scale wavelet transform and energy entropy in feature extraction. The constructed wavelet energy entropy feature vector retains both frequency domain information and distribution characteristics, reducing model learning difficulty and data dependence. The network adopts a lightweight design with deep separable convolution and channel attention mechanism, achieving 8-millisecond real-time response while ensuring classification accuracy, solving the problems of high computational cost and difficulty in real-time deployment of traditional deep learning models. Combined with the DTW algorithm to handle time distortion, the localization error does not exceed 0.1 milliseconds, far exceeding the accuracy of traditional methods. Robustness is enhanced by preprocessing denoising, noise injection during training, and wavelet domain-specific loss functions, improving generalization ability under complex disturbance scenarios and solving the pain point of weak generalization of traditional threshold methods. This application forms a complete technical closed loop of "detection-classification-localization", balancing detection speed and accuracy, and can be directly applied to power system power quality analysis and protection control scenarios.
[0089] Example 2:
[0090] This embodiment provides a distribution network transformer area voltage transient event detection system based on wavelet transform and lightweight deep learning, including: a memory and a processor;
[0091] The memory is used to store computer programs;
[0092] The processor is configured to invoke the computer program to execute the method as described in Embodiment 1.
[0093] Example 3:
[0094] This embodiment provides a computer-readable storage medium storing a computer program. When the computer program is run on an electronic device, it causes the electronic device to perform the method described in Embodiment 1.
[0095] Example 4:
[0096] This embodiment provides a computer program product, including a computer program that, when run on an electronic device, causes the electronic device to perform the method described in Embodiment 1.
[0097] The specific implementation of the system, electronic device, computer-readable storage medium, and computer program product provided in this application can be referred to the specific embodiments of the above methods, and will not be repeated here.
[0098] Obviously, those skilled in the art should understand that the various units or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device, or fabricating them separately as individual integrated circuit modules, or fabricating multiple modules or steps into a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.
[0099] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A method for detecting voltage transient events in distribution network transformer areas based on wavelet transform and lightweight deep learning, characterized in that, include: Step S1: Preprocess and perform multi-scale wavelet transform on the three-phase power grid voltage signal of the distribution network area to extract the wavelet energy entropy feature vector; Step S2: Construct and train a lightweight deep learning classification network, using the wavelet energy entropy feature vector as input, and output the probability corresponding to each type of voltage transient event to complete the voltage transient event type determination. Step S3: For the voltage over-limit transient events that are determined to be voltage over-limit, calibrate their start and end times based on the dynamic time warping algorithm; Step S4: Output the detection results, including the voltage transient event type and its confidence level, and the start and end times of the voltage over-limit transient event; where the confidence level of the voltage transient event type refers to the probability that the lightweight deep learning classification network determines the type of the current event.
2. The method according to claim 1, characterized in that, Step S1 specifically includes: S11: Acquire three-phase power grid voltage signals and perform noise reduction and normalization preprocessing; S12: Perform multi-level discrete wavelet decomposition on the preprocessed signal to obtain the high-frequency detail coefficients and low-frequency approximation coefficients of each level; S13: Calculate the energy and wavelet energy entropy of each frequency band; S14: Combine the energy and energy entropy of each frequency band to form a multi-scale wavelet energy entropy feature vector.
3. The method according to claim 2, characterized in that, In step S13, the energy calculation formula for each frequency band is as follows: ； in, For the first Energy of each frequency band Representing the The first frequency band Wavelet coefficients, For the first The coefficient length of each frequency band; The formula for calculating wavelet energy entropy is: in, For wavelet energy entropy, For the first The energy of each frequency band accounts for a portion of the total energy. The ratio, i.e. ; is the number of frequency bands, where It equals the number of layers in the discrete wavelet decomposition.
4. The method according to claim 1, characterized in that, The lightweight deep learning classification network comprises, in sequence: an input layer, a structured feature remodeling layer, a multi-scale deep separable convolutional fusion module, a joint attention module, a global average pooling layer, and an output layer. The structured feature renormalization layer is used to rearrange the wavelet energy entropy feature vector according to phase, frequency band and statistics to form a three-dimensional feature tensor, and generate at least one set of interphase differential feature channels in the three-dimensional feature tensor to characterize the three-phase coupling relationship; The multi-scale depth-separable convolutional fusion module includes at least two parallel depth-separable convolutional branches. Each depth-separable convolutional branch includes a depth-separable convolutional layer, a batch normalization layer, and a ReLU activation function connected in sequence. Different depth-separable convolutional branches use different kernel sizes and / or different dilatation rates to extract transient discriminative features at different scales. The outputs of each depthwise separable convolutional branch are fused through pointwise convolution to obtain a fused feature map; the fused feature map is then superimposed on the input features of the multi-scale depthwise separable convolutional fusion module through residual connections. The joint attention module includes a channel attention submodule and a frequency band attention submodule. The channel attention submodule generates channel weights based on the channel statistical descriptors obtained by global average pooling, and the frequency band attention submodule generates frequency band weights based on the statistical descriptors of the frequency band dimension. The channel weights and frequency band weights are multiplied by the corresponding dimensions of the input features of the joint attention module to strengthen key transient features and suppress redundant features, thereby obtaining a weighted three-dimensional feature tensor. The output layer includes a confidence calibration unit and a Softmax classifier, which output the probability of various voltage transient events.
5. The method according to claim 1, characterized in that, The training process of the lightweight deep learning classification network uses a wavelet domain-specific loss function, and noisy data is injected into the training dataset.
6. The method according to claim 1, characterized in that, Step S3 specifically includes: S31: When the lightweight deep learning classification network determines that the event is a voltage over-limit transient event, extract the high-frequency wavelet coefficient sequence of the event period; S32: Perform dynamic time warping calculation on the high-frequency wavelet coefficient sequence and the preset typical transient disturbance template, and output the dynamic time warping cost matrix; S33: Based on the dynamic time warping cost matrix, find the optimal path and align the distortion points of the high-frequency wavelet coefficient sequence with the preset typical transient disturbance template through the optimal path; S34: Determine the start and end points of voltage over-limit transient events based on aligned distortion points.
7. The method according to claim 1, characterized in that, The preset typical transient disturbance templates in step S32 include three core templates: voltage sag, voltage swell, and instantaneous interruption. Each template is generated by statistical modeling of high-frequency wavelet coefficient sequences of no less than 1,000 sets of standard transient event samples. The templates support online updates. When a new type of transient limit-breaking event is detected, its high-frequency wavelet coefficient features are automatically extracted and added to the template library.
8. A distribution network transformer area voltage transient event detection system based on wavelet transform and lightweight deep learning. The feature is that it includes: a memory and a processor; The memory is used to store computer programs; The processor is configured to invoke the computer program to perform the method as described in any one of claims 1 to 7.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed on an electronic device, causes the electronic device to perform the method as described in any one of claims 1 to 7.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is run on an electronic device, it causes the electronic device to perform the method as described in any one of claims 1 to 7.