CN-122020369-A - Power distribution network area voltage transient event detection method and system based on wavelet transformation and lightweight deep learning

CN122020369ACN 122020369 ACN122020369 ACN 122020369ACN-122020369-A

Abstract

The application provides a power distribution network station voltage transient event detection method and a power distribution network station voltage transient event detection system based on wavelet transformation and lightweight deep learning, wherein the method comprises the steps of preprocessing three-phase power grid voltage signals of the power distribution network station and performing multi-scale wavelet transformation to extract wavelet energy entropy feature vectors; the lightweight deep learning classification network is constructed and trained, the wavelet energy entropy feature vector is taken as input, the probability corresponding to various voltage transient events is output, the voltage transient event type judgment is completed, the starting and stopping time of the judged voltage out-of-limit transient event is calibrated based on a dynamic time warping algorithm, the detection result is output, the detection result comprises the voltage transient event type, the confidence level of the voltage transient event type and the starting and stopping time of the voltage out-of-limit transient event, and the confidence level of the voltage transient event type refers to the probability corresponding to the type of the lightweight deep learning classification network judged to be the current event. The method can improve the real-time performance, the accuracy and the robustness of the voltage transient detection.

Inventors

ZHENG JUNFENG
PAN SHUYANG
CHEN NAN
HUANG JIYUAN
LI LIN
ZHONG JIACHEN
CHEN BAIYUAN
Zuo Xinya
YI JIANG
FU XIANG

Assignees

国网湖南省电力有限公司长沙供电分公司
国网湖南省电力有限公司
国家电网有限公司

Dates

Publication Date: 20260512
Application Date: 20260119

Claims (10)

1. The utility model provides a power distribution network district voltage transient event detection method based on wavelet transformation and lightweight deep learning, which is characterized by comprising the following steps: step S1, preprocessing and multi-scale wavelet transformation are carried out on three-phase power grid voltage signals of a power distribution network area, and wavelet energy entropy feature vectors are extracted; S2, constructing and training a lightweight deep learning classification network, taking the wavelet energy entropy feature vector as input, outputting probabilities corresponding to various voltage transient events, and finishing voltage transient event type judgment; S3, calibrating the start-stop time of the determined voltage out-of-limit transient event based on a dynamic time warping algorithm; and S4, outputting a detection result, wherein the detection result comprises a voltage transient event type, a confidence coefficient thereof and a start-stop time of a voltage out-of-limit transient event, and the confidence coefficient of the voltage transient event type refers to the probability corresponding to the type to which the lightweight deep learning classification network judges the current event belongs.
2. The method according to claim 1, wherein the step S1 specifically comprises: S11, collecting a three-phase power grid voltage signal, and carrying out denoising and normalization pretreatment; S12, carrying out multi-layer discrete wavelet decomposition on the preprocessed signals to obtain high-frequency detail coefficients and low-frequency approximation coefficients of each layer; S13, calculating energy of each frequency band and wavelet energy entropy; And S14, combining energy of each frequency band with energy entropy to form a multi-scale wavelet energy entropy feature vector.
3. The method according to claim 2, wherein in the step S13, each band energy calculation formula is: ; Wherein, the Is the first The energy of the individual frequency bands is used, Represents the first Under the frequency band The number of coefficients of the wavelet is the number of coefficients, Is the first Coefficient lengths of the individual frequency bands; The wavelet energy entropy calculation formula is: Wherein, the For the wavelet energy entropy, Is the first The energy of the individual frequency bands occupying the total energy Ratio of (2), i.e ; Is the number of frequency bands, wherein Equal to the number of layers of the discrete wavelet decomposition.
4. The method of claim 1, wherein the lightweight deep learning classification network comprises an input layer, a structured feature reformation layer, a multi-scale deep separable convolution fusion module, a joint attention module, a global average pooling layer, and an output layer, connected in sequence; The structural feature reforming layer is used for rearranging the wavelet energy entropy feature vectors according to the phase, the frequency band and the statistic to form a three-dimensional feature tensor, and generating at least one group of phase-to-phase difference feature channels in the three-dimensional feature tensor to represent a three-phase coupling relation; The multi-scale depth separable convolution fusion module comprises at least two parallel depth separable convolution branches, wherein each depth separable convolution branch comprises a depth separable convolution layer, a batch normalization layer and a ReLU activation function which are sequentially connected, different depth separable convolution branches adopt different convolution kernel sizes and/or different void ratios to extract transient discrimination features of different scales, and the outputs of the depth separable convolution branches are fused through point-by-point convolution to obtain a fusion feature map; The joint attention module comprises a channel attention sub-module and a frequency band attention sub-module, wherein the channel attention sub-module generates channel weights based on channel statistical descriptors obtained by global average pooling, the frequency band attention sub-module generates frequency band weights based on statistical description of frequency band dimensions, and the channel weights and the frequency band weights are multiplied with corresponding dimensions of input features of the joint attention module respectively so as to strengthen key transient features and inhibit redundant features and obtain weighted three-dimensional feature tensors; The output layer comprises a confidence calibration unit and a Softmax classifier, and outputs probabilities of various voltage transient events.
5. The method of claim 1, wherein the training process of the lightweight deep learning classification network employs a wavelet domain specific loss function and noise data is injected in a training dataset.
6. The method according to claim 1, wherein the step S3 specifically includes: s31, when the lightweight deep learning classification network judges that the voltage is out of limit transient event, extracting a high-frequency wavelet coefficient sequence of an event period; S32, carrying out dynamic time warping calculation on the high-frequency wavelet coefficient sequence and a preset typical transient disturbance template, and outputting a dynamic time warping cost matrix; s33, searching an optimal path based on a dynamic time warping cost matrix, and aligning the high-frequency wavelet coefficient sequence with a distortion point of a preset typical transient disturbance template through the optimal path; s34, determining a starting point and an ending point of the voltage out-of-limit transient event based on the aligned distortion points.
7. The method according to claim 1, wherein the preset typical transient disturbance templates in the step S32 include three types of core templates including a voltage dip, and a transient interrupt, each type of template is generated by statistical modeling of a high-frequency wavelet coefficient sequence of not less than 1000 sets of standard transient event samples, and the templates support online updating, and when a new type of transient out-of-limit event is detected, the high-frequency wavelet coefficient features thereof are automatically extracted and supplemented to a template library.
8. A power distribution network station voltage transient event detection system based on wavelet transformation and lightweight deep learning is characterized by comprising a memory and a processor; the memory is used for storing a computer program; The processor for invoking the computer program to perform the method of any of claims 1 to 7.
9. A computer readable storage medium, characterized in that the computer readable storage medium has stored therein a computer program which, when run on an electronic device, causes the electronic device to implement the method of any one of claims 1 to 7.
10. A computer program product comprising a computer program which, when run on an electronic device, causes the electronic device to carry out the method of any one of claims 1 to 7.

Description

Power distribution network area voltage transient event detection method and system based on wavelet transformation and lightweight deep learning Technical Field The invention relates to the technical field of power quality analysis and protection control of power systems, in particular to a power distribution network station voltage transient event detection method and system based on wavelet transformation and lightweight deep learning. Background With the frequent occurrence of transient out-of-limit events such as voltage sag and transient rise of a power grid along with the generation of new energy, the large number of accesses of a distributed power supply and a nonlinear load in a power system, a high requirement is provided for a rapid and accurate detection method. The traditional voltage out-of-limit detection method is mostly based on voltage effective value calculation or second-order generalized integrator to carry out fundamental wave extraction and phase synchronization. However, the effective value algorithm has calculation delay, so that instantaneous mutation is difficult to capture, and the integrator methods such as SOGI/DSOGI have response lag and estimation deviation problems under voltage mutation and harmonic interference. The prior improved scheme is as disclosed in patent CN120879642A, adopts wavelet transformation to locate voltage mutation time, but the core of the improved scheme still depends on direct comparison of wavelet coefficients and fixed thresholds, lacks intelligent learning ability and has weak generalization ability to complex disturbance. Another patent CN120855539a proposes a system-level voltage out-of-limit abatement scheme, but its abatement performance is limited by the speed and accuracy of the underlying detection technique. Therefore, a new method for detecting voltage transient out-of-limit is needed, which can achieve real-time performance, accuracy and robustness. Disclosure of Invention The invention aims to provide a power distribution network station voltage transient event detection method and system based on wavelet transformation and lightweight deep learning, which are used for solving the problems of low response speed and insufficient precision under complex disturbance in the prior detection technology. In order to achieve the above purpose, the invention adopts the following technical scheme: A power distribution network area voltage transient event detection method based on wavelet transformation and lightweight deep learning comprises the following steps: step S1, preprocessing and multi-scale wavelet transformation are carried out on three-phase power grid voltage signals of a power distribution network area, and wavelet energy entropy feature vectors are extracted; S2, constructing and training a lightweight deep learning classification network, taking the wavelet energy entropy feature vector as input, outputting probabilities of various voltage transient events (such as dip, sag and the like), and finishing voltage transient event type judgment; S3, calibrating the start-stop time of the determined voltage out-of-limit transient event based on a dynamic time warping algorithm; and S4, outputting a detection result, wherein the detection result comprises a voltage transient event type, a confidence coefficient thereof and a start-stop time of a voltage out-of-limit transient event, and the confidence coefficient of the voltage transient event type refers to the probability corresponding to the type to which the lightweight deep learning classification network judges the current event belongs. In one possible implementation manner, the step S1 specifically includes: S11, collecting a three-phase power grid voltage signal, and carrying out denoising and normalization pretreatment; S12, carrying out multi-layer discrete wavelet decomposition on the preprocessed signals to obtain high-frequency detail coefficients and low-frequency approximation coefficients of each layer; S13, calculating energy of each frequency band and wavelet energy entropy; And S14, combining energy of each frequency band with energy entropy to form a multi-scale wavelet energy entropy feature vector. In one possible implementation manner, in the step S13, the energy calculation formula of each frequency band is: ; Wherein, the Is the firstThe energy of the individual frequency bands is used,Represents the firstUnder the frequency bandThe number of coefficients of the wavelet is the number of coefficients,Is the firstCoefficient lengths of the individual frequency bands; The wavelet energy entropy calculation formula is: Wherein, the For the wavelet energy entropy,Is the firstThe energy of the individual frequency bands occupying the total energyRatio of (2), i.e;Is the number of frequency bands, whereinEqual to the number of layers of the discrete wavelet decomposition. In one possible implementation manner, the step S2 specifically includes: in one possible implemen