Method and system for fast classification of power quality disturbances in photovoltaic grid-connected systems

Abstract

The application discloses a kind of photovoltaic grid-connected system electric energy quality disturbance dimensionality reduction fast classification method and system.The method is fused into one-dimensional CTM characteristic waveform by Clark transformation to three-phase voltage signal, while retaining asymmetric and unbalanced disturbance characteristics completely, substantially reduce data dimension, significantly reduce memory occupation and shorten the time of calculation consumption.Subsequently, multi-layer maximum overlap discrete wavelet transform is used, and the shift sensitivity and boundary effect of traditional transform are eliminated using the overlapping filter mechanism, and the multi-scale time-frequency characteristics of highly non-stationary are accurately extracted.Finally, adaptive classification is carried out through time convolution network, relying on full convolution architecture and causal convolution to realize large-scale parallel computing and timing logic constraint, combined with dilated convolution and residual module to capture high-frequency transient distortion and low-frequency sustained fluctuation simultaneously.The application has strong noise robustness and can still maintain high accuracy in strong interference environment, fully meets the high robustness requirement of new power system for real-time monitoring of electric energy quality.

Description

Technical Field

[0001] This invention relates to the field of power system power quality monitoring and intelligent operation technology of distribution networks, specifically to a method and system for rapid classification of power quality disturbances in photovoltaic grid-connected systems. Background Technology

[0002] With the acceleration of the global energy transition, new energy power generation technologies, represented by distributed photovoltaic (PV) power, have developed rapidly, and building a modern new power system with a high proportion of new energy grid connection has become an inevitable trend. However, PV power generation systems have significant intermittency and volatility, and their grid connection points extensively use nonlinear equipment such as power electronic inverters. Complex inverter interaction coupling effects and rapid changes in grid source load lead to frequent complex power quality disturbance events at PV grid connection points, such as voltage sags, voltage swells, harmonic pollution, flicker, transient oscillations, voltage gaps, and voltage spikes. These disturbances may cause relay protection devices to malfunction, damage precision electronic equipment, and in severe cases, even trigger chain reactions, threatening the stable operation of the entire PV grid-connected system. Therefore, how to identify and classify various power quality disturbance events in real time and accurately is a core issue in ensuring the reliable power supply of modern power grids.

[0003] In actual operation of power distribution networks, due to the randomness of environmental factors and the complexity of inverter control strategies, power quality disturbances at the grid connection point often exhibit significant asymmetric and unbalanced characteristics. Therefore, synchronous real-time monitoring of three-phase voltage signals is essential. However, traditional comprehensive analysis methods face the following bottlenecks when directly processing massive three-phase signals: High dimensionality and large computational load. Simultaneously processing data from three phases requires three times the storage space and enormous computing resources, resulting in extremely high memory consumption during data transmission, storage, and model training, making it difficult to meet the requirements of real-time monitoring.

[0004] Limitations of traditional time-frequency analysis methods. Fast Fourier Transform (FFT) suffers from spectral leakage and loss of temporal information; the fixed window length of Short-Time Fourier Transform (SFT) makes it unsuitable for non-stationary transient signals; while traditional Discrete Wavelet Transform (DWT) possesses time-frequency localization capabilities, it is often affected by shift sensitivity and boundary effects caused by downsampling when processing photovoltaic grid-connected signals accompanied by strong switching noise and highly non-stationary transient components. Insufficient thresholding methods and traditional intelligent classifiers. Detection techniques based on system parameter thresholds require a stringent balance between false alarm rate and accuracy, and are highly sensitive to harmonic components and sampling frequency; traditional machine learning classifiers (such as support vector machines and decision trees) often face computational bottlenecks when extracting features from large-scale nonlinear data, while conventional recurrent neural networks are prone to gradient vanishing or exploding problems when processing long sequences of data spanning multiple fundamental frequency periods, and do not support large-scale parallel computing, making them unable to adaptively and efficiently mine weak high-frequency transient features.

[0005] In summary, existing classification methods often face bottlenecks when processing massive three-phase imbalance signals in high-proportion distributed photovoltaic grid-connected systems, including heavy computational load, poor noise immunity, and difficulty in capturing weak transient distortions. Therefore, there is an urgent need for an intelligent solution that can significantly reduce signal dimensionality, optimize memory utilization, effectively suppress edge effects, and adaptively extract multi-scale temporal features under strong switching noise interference, thereby achieving accurate classification. Summary of the Invention

[0006] This invention provides a method for rapid classification of power quality disturbances in photovoltaic grid-connected systems with reduced dimensions. It solves the technical problems of high-dimensional power quality disturbance signals, large computational load, difficulty in capturing non-stationary transient features, and poor noise resistance of traditional methods caused by high proportion of distributed photovoltaic grid connection.

[0007] This invention is achieved through the following technical solution:

[0008] Firstly, this application provides a method for rapid classification of power quality disturbances in photovoltaic grid-connected systems using dimensionality reduction, comprising the following steps:

[0009] Collect the three-phase voltage signal of the photovoltaic grid connection point, and normalize the three-phase voltage signal to obtain the three-phase voltage signal in per-unit form;

[0010] The three-phase voltage signal in per-unit form is subjected to Clarke transform to decouple it and obtain α component, β component and zero-sequence component. The α component and the β component are multiplied and then the zero-sequence component is superimposed to obtain a one-dimensional CTM characteristic waveform.

[0011] The one-dimensional CTM feature waveform is decomposed by multi-level maximum overlap discrete wavelet transform to extract multiple high-frequency detail components and one low-frequency approximation component. The components are then concatenated along the channel dimension to form a multi-channel time series matrix.

[0012] The multi-channel time series matrix is ​​input into the temporal convolutional network model, and the temporal dependency features are extracted through causal convolution and dilated convolution to obtain the power quality disturbance type of the corresponding photovoltaic grid-connected system.

[0013] A further optimization scheme involves normalizing the three-phase voltage signals, specifically including:

[0014] The fundamental frequency of the power grid is set to the rated power frequency, and the three-phase voltage signal is intercepted in real time through a multi-cycle buffer sliding window so that each sample contains a preset number of discrete sampling points;

[0015] The amplitude of each sampling point is normalized and converted into a per-unit signal without amplitude dimension.

[0016] A further optimized solution is that the specific formula for the Clarke transform is:

[0017] ;

[0018] In the formula, For the α component, For the β component, For zero-order components, This represents the instantaneous value of phase A voltage. This is the instantaneous value of phase B voltage. This is the instantaneous value of the C-phase voltage.

[0019] A further optimization is that the calculation formula for the one-dimensional CTM feature waveform is:

[0020] ;

[0021] In the formula, This is a one-dimensional CTM characteristic waveform.

[0022] A further optimization scheme is as follows: in the step of performing multi-level maximum overlap discrete wavelet transform decomposition on the one-dimensional CTM feature waveform, extracting multiple high-frequency detail components and one low-frequency approximate component, and splicing each component along the channel dimension to form a multi-channel time series matrix, high-pass and low-pass filters are used alternately for filtering, and overlap is maintained between adjacent sub-bands to filter out the shift sensitivity and boundary effects caused by traditional discrete wavelet transform downsampling.

[0023] A further optimization scheme is that the temporal convolutional network model includes multi-level cascaded residual modules. Each residual module contains at least two levels of cascaded dilated causal convolutions. The convolution kernel size is a preset odd number. The number of filters in each layer increases exponentially with the network depth, and the dilation factor increases exponentially with the number of layers to cover the temporal span of the sampling points.

[0024] A further optimization is that the training process of the temporal convolutional network model includes:

[0025] Set the initial learning rate, batch size for a single training session, and maximum number of iterations;

[0026] During training, the learning rate is decayed in segments at preset iteration intervals to drive the model to converge to the global optimum.

[0027] A further optimization scheme is as follows: after each dilated causal convolution operation, the residual module sequentially connects a weight normalization layer, a nonlinear activation layer, and a regularization layer, and adds the input and output of the residual module element by element through skip connections; when the number of input channels is inconsistent with the number of output channels, a convolutional layer is introduced on the skip connection branch to adjust the feature dimension so that the number of input and output channels matches.

[0028] A further optimization is that the operation logic of the causal convolution follows the following formula:

[0029] ;

[0030] In the formula, Let x be the output feature at the current position s, and x be the input sequence. Here, k is the kernel size, and k is the filter. This indicates that the current position s in the input signal is to the left. The value at the distance, d is the dilation rate, and i is the index of the current convolution kernel position.

[0031] Secondly, this application provides a dimensionality-reduced rapid classification system for power quality disturbances in photovoltaic grid-connected systems, comprising:

[0032] The signal acquisition module is used to capture the three-phase voltage signal at the photovoltaic grid connection point in real time and perform normalization processing to obtain the three-phase voltage signal in per-unit form;

[0033] The feature extraction module is used to perform Clarke transform on the three-phase voltage signal in per-unit form to obtain a one-dimensional CTM feature waveform, and then decompose it through multi-layer maximum overlap discrete wavelet transform to generate a multi-channel time sequence matrix.

[0034] The classification module is used to load the trained temporal convolutional network model, receive the multi-channel time series matrix, and output the corresponding power quality perturbation type.

[0035] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0036] By using Clarke transform fusion dimensionality reduction, the data dimensionality is significantly reduced while fully preserving the core characteristics of three-phase asymmetry and unbalanced disturbances. Calculations show that up to 66.7% reduction in data dimensionality can be achieved, significantly reducing memory usage and shortening computation time.

[0037] By employing an overlap filtering mechanism based on multi-level maximum overlap discrete wavelet transform, the shift sensitivity of traditional wavelet transform is eliminated at its source, and boundary effects are suppressed, thus accurately characterizing the highly non-stationary multi-scale time-frequency characteristics unique to photovoltaic grid connection.

[0038] The fully convolutional architecture of temporal convolutional networks supports massively parallel computing to improve processing efficiency. Its causal convolutional structure strictly follows temporal logic, effectively eliminating the risk of future information leakage in online analysis.

[0039] Dilated convolution achieves exponential expansion of the receptive field with low computational cost. Combined with hierarchical feature extraction and regularization design of the residual module, the model can simultaneously capture microsecond-level high-frequency transient distortions and cross-cycle low-frequency continuous fluctuations.

[0040] This method exhibits strong noise resistance and robustness under strong electromagnetic interference and heavy noise environments, demonstrating significant engineering practical value and fully meeting the technical requirements of new power systems for high-precision and high-robustness real-time power quality monitoring. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0042] Figure 1 A flowchart illustrating a method for rapid classification of power quality disturbances in a photovoltaic grid-connected system provided in an embodiment of this application;

[0043] Figure 2 A schematic diagram illustrating the principle of the discrete wavelet transform decomposition process provided in the embodiments of this application;

[0044] Figure 3 This is an overall flowchart of CTM-MODWT-TCN signal processing and classification provided in the embodiments of this application;

[0045] Figure 4 This is a schematic diagram of the internal structure of the TCN core residual module provided in an embodiment of this application;

[0046] Figure 5 A comparative schematic diagram showing the original asymmetrical three-phase voltage waveform provided in the embodiments of this application and the one-dimensional CTM feature signal generated after Clarke transformation;

[0047] Figure 6 Convergence curves showing the increase in accuracy and the decrease in loss function of the TCN model provided in the embodiments of this application during the training and validation process;

[0048] Figure 7 The convergence curve of the TCN model provided in the embodiments of this application during the training and validation process;

[0049] Figure 8 The confusion matrix diagram of classification test performance under different noisy conditions provided in the embodiments of this application. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0051] First, some of the technical terms used in this application will be explained to help those skilled in the art understand this application.

[0052] CTM: Clark Transform Modulation;

[0053] MODWT: Maximal Overlap Discrete Wavelet Transform;

[0054] TCN: Temporal Convolutional Network;

[0055] SCTM: Single Channel CTM, a single-channel Clarke transform mode;

[0056] IEEE-1159: IEEE Recommended Practice for Monitoring Electric Power Quality.

[0057] Firstly, such as Figure 1 As shown, this application provides a method for rapid classification of power quality disturbances in photovoltaic grid-connected systems with dimensionality reduction, including the following steps:

[0058] Step S1: Collect the three-phase voltage signal at the photovoltaic grid connection point, and normalize the three-phase voltage signal to obtain the three-phase voltage signal in per-unit form;

[0059] Step S2: Perform Clarke transform on the three-phase voltage signal in per-unit form to decouple and obtain α component, β component and zero-sequence component. Multiply the α component and the β component and then superimpose the zero-sequence component to obtain a one-dimensional CTM characteristic waveform.

[0060] Step S3: Perform multi-level maximum overlap discrete wavelet transform decomposition on the one-dimensional CTM feature waveform, extract multiple high-frequency detail components and one low-frequency approximation component, and splice the components along the channel dimension to form a multi-channel time series matrix;

[0061] Step S4: Input the multi-channel time series matrix into the temporal convolutional network model, extract the temporal dependency features through causal convolution and dilated convolution, and obtain the power quality disturbance type of the corresponding photovoltaic grid-connected system.

[0062] Among them, dilated convolution is the standard term in the field of deep learning, "dilated convolution";

[0063] This embodiment utilizes the Clark transform fusion dimensionality reduction technique to achieve a significant reduction in data dimensionality while fully preserving the core characteristics of asymmetric and unbalanced disturbances in a three-phase system, thereby significantly reducing computational load and memory usage.

[0064] Multi-level maximum overlap discrete wavelet transform decomposition is adopted. By using the overlap mechanism between adjacent sub-bands, the shift sensitivity and boundary effect caused by traditional discrete wavelet transform downsampling are effectively eliminated, and multi-scale time-frequency features including high-frequency transients and low-frequency trends are accurately extracted.

[0065] By introducing a temporal convolutional network model, and utilizing causal convolution and dilated convolution mechanisms, the receptive field is expanded exponentially while preventing the leakage of future information. It adaptively mines weak high-frequency transient features and long-distance temporal dependencies, achieving extremely high average classification accuracy in environments with strong electromagnetic interference. It combines high precision with strong noise robustness.

[0066] In one embodiment, step S1: Acquire the three-phase voltage signal at the photovoltaic grid connection point, and normalize the three-phase voltage signal to obtain the three-phase voltage signal in per-unit form. This specifically includes the following steps:

[0067] Step S11: Use a preset duration cyclic buffer to capture the three-phase voltage waveform of the photovoltaic grid connection point in real time using a sliding window, convert the captured continuous analog signal into a digital sequence, and obtain the original three-phase voltage sampling data composed of discrete sampling points.

[0068] The original three-phase voltage signal originates from the complex electrical response generated by the intermittent and fluctuating environmental conditions at the photovoltaic grid connection point, as well as the inherent nonlinear interactive coupling of the inverter. This results in the signal having significant non-stationary and unbalanced characteristics, thus requiring specific signal processing methods to achieve accurate feature extraction and classification.

[0069] Specifically, the preset duration of the circular buffer is a 6-cycle buffer.

[0070] Specifically, the fundamental frequency of the power system is set to 50Hz, and the sampling frequency of the signal is controlled to 6400Hz, so that the duration of each sample captured by the sliding window is 0.2s, thereby ensuring that each sample strictly contains 1280 discrete sampling points.

[0071] Step S12: Perform per-unit value conversion on the original three-phase voltage sampling data composed of the discrete sampling points one by one to generate a standard voltage signal in per-unit value form. The conversion formula is as follows:

[0072] ;

[0073] In the formula, This represents the converted per-unit value. This indicates the actual voltage amplitude collected. This indicates the amplitude of the system reference voltage.

[0074] Step S13: Perform three-phase reconstruction on the standard voltage signal in per-unit form to obtain the three-phase voltage signal in per-unit form, and mark it as... , , These are the per-unit voltage signals for phases A, B, and C, respectively.

[0075] This embodiment ensures the temporal consistency of input data by setting a fixed sampling frequency and sliding window duration, providing a standardized data foundation for subsequent feature extraction. At the same time, through amplitude normalization processing, the dimensional differences caused by different voltage levels or load fluctuations are eliminated, significantly improving the model's generalization adaptability to different operating conditions.

[0076] In one embodiment, step S2: performing a Clarke transform on the per-unit three-phase voltage signal to decouple and obtain an α component, a β component, and a zero-sequence component; multiplying the α component and the β component and then superimposing the zero-sequence component to obtain a one-dimensional CTM characteristic waveform, specifically including the following steps:

[0077] Step S21: Map the three-phase voltage signal in per-unit form to the α-β-0 stationary reference coordinate system, and decouple to obtain orthogonal components, including α component, β component and zero-sequence component characterizing unbalance.

[0078] Specifically, the three-phase system is projected onto coordinates using the Clarke transformation matrix, and the transformation formula is as follows:

[0079] ;

[0080] In the formula, For the α component, For the β component, For zero-order components, This represents the instantaneous value of phase A voltage. This is the instantaneous value of phase B voltage. This is the instantaneous value of the C-phase voltage.

[0081] Step S22: Based on orthogonal components and The phase correlation is multiplied to construct an intermediate variable containing second harmonic characteristics, and then combined with the zero-sequence component. The one-dimensional Clarke transform modal signal is extracted, and its operation logic is as follows:

[0082] ;

[0083] In the formula, It is a one-dimensional Clarke transform modal signal.

[0084] This embodiment utilizes the Clarke transform fusion dimensionality reduction mechanism to fully preserve the core characteristics of asymmetric and unbalanced disturbances in the three-phase system. It uses a single-channel SCTM signal to fully carry the electrical characteristic information of the original three-phase system, achieving a data dimensionality reduction of up to 66.7%. This effectively optimizes memory usage and significantly shortens processing latency, providing efficient feature input for subsequent time-frequency analysis and deep learning classification.

[0085] In one embodiment, step S3: performing multi-level maximum overlap discrete wavelet transform decomposition on the one-dimensional CTM feature waveform, extracting multiple high-frequency detail components and one low-frequency approximation component, and concatenating the components along the channel dimension to form a multi-channel time series matrix, specifically including the following steps:

[0086] Step S31: Perform multi-level maximum overlap discrete wavelet transform (MODWT) decomposition on the one-dimensional CTM feature waveform. This is achieved through a cascaded high-pass filter. With low-pass filter Alternating filtering is performed, maintaining overlap between adjacent sub-bands, to filter out the shift sensitivity and boundary effects caused by traditional discrete wavelet transform downsampling, thus obtaining the multi-scale decomposition results. The filtering process of the j-th layer can be expressed as:

[0087] ;

[0088] In the formula, This represents the detail coefficient at the k-th position in the j-th layer. Indicates the input signal. This represents the filter coefficients of the j-th layer.

[0089] Step S32: Extract multiple detail components for capturing high-frequency transients or switching noise and one low-frequency approximation component for characterizing the basic contour of the signal from the multi-scale decomposition results.

[0090] After multi-level decomposition, the output contains multiple high-frequency detail components D1, D2, ..., DJ and one low-frequency approximation component AJ, where J is the decomposition level.

[0091] Specifically, after 5-level decomposition, the output includes 5 high-frequency detail components D1, D2, D3, D4, and D5 and 1 low-frequency approximation component A5, each component corresponding to time-frequency features at different scales.

[0092] Step S33: Perform matrix concatenation of the multiple high-frequency detail components and one low-frequency approximation component along the channel dimension to construct a multi-channel timing matrix.

[0093] Specifically, the components are horizontally concatenated to form a two-dimensional matrix containing J+1 feature channels, which serves as the direct input feature of the temporal convolutional network model.

[0094] Specifically, the five high-frequency detail components (D1~D5) and one low-frequency approximation component (A5) are matrix-joined along the channel dimension to form a two-dimensional matrix containing six feature channels.

[0095] This embodiment fundamentally eliminates the shift sensitivity and boundary effects caused by traditional discrete wavelet transform downsampling through the overlap filtering mechanism of maximum overlap discrete wavelet transform. It can more accurately characterize the highly non-stationary multi-scale time-frequency characteristics unique to photovoltaic grid connection, providing a highly discriminative input representation for subsequent classification models.

[0096] In one embodiment, step S4: Input the multi-channel time series matrix into a temporal convolutional network model, extract temporal dependency features through causal convolution and dilated convolution, and obtain the power quality disturbance type of the corresponding photovoltaic grid-connected system. This specifically includes the following steps:

[0097] Step S41: Input the multi-channel temporal matrix into the bottom causal convolutional layer of the temporal convolutional network, and use the unidirectional connection rule to restrict the feature propagation direction to generate basic features that preserve the temporal order.

[0098] Specifically, the operational logic of causal convolution follows the formula below, ensuring that the output at any time t depends only on the input before t:

[0099] ;

[0100] In the formula, x is the input sequence; For filters; k is the kernel size; This indicates that the current position s in the input signal is to the left. The value at the distance is given by d, where d is the dilation rate and i is the position index of the current convolution kernel. This discrete dilated convolution mechanism introduces a dilated convolution mechanism, which performs strided sampling on the input multi-channel sequence by setting a dilation factor d, enabling higher-level nodes of the network to connect to more distant historical inputs in a skip manner, exponentially expanding the receptive field without significantly increasing the network depth and computational load.

[0101] Step S42: Apply an exponentially dilated convolution operation to the basic features to expand the feature perception range to cover long-distance temporal dependencies and obtain deep features containing global temporal information.

[0102] Specifically, the temporal convolutional network model is constructed by cascading three core residual modules. The dilation factor increases exponentially with the number of network layers, with the dilation rates of layers 1 to 3 set to 1, 2, and 4, respectively. The number of filters in each layer doubles with the network depth, at 32, 64, and 128, respectively. Combined with a convolutional kernel size of 5, this allows the receptive field to fully cover the temporal span of 1280 sampling points without significantly increasing the network depth and computational load.

[0103] Step S43: Input deep features into the cascaded residual module, and obtain classification features with strong noise resistance through nonlinear transformation and feature skipping fusion.

[0104] Each residual module contains two cascaded dilated causal convolutions, processed through a normalization and activation mapping chain: after each dilated causal convolution operation, a weight normalization layer is sequentially connected to accelerate convergence, a rectified linear unit activation layer is introduced to introduce nonlinearity, and a dropout layer is used to suppress overfitting; when the number of input and output channels is inconsistent, the feature dimension of the skip connections is adjusted through 1×1 convolution to achieve smooth cross-layer transfer of low-level features.

[0105] Step S44: Perform a fully connected mapping on the classification features to obtain the power quality disturbance type of the corresponding photovoltaic grid-connected system.

[0106] Specifically, the power quality disturbance types cover eight typical power quality disturbance signals caused by inverter grid-connection / off-grid switching, maximum power point tracking regulation, and sudden weather changes, including:

[0107] (1) Normal signal C1;

[0108] (2) Voltage spurt disturbance signal C2;

[0109] (3) Voltage sag disturbance signal C3;

[0110] (4) Harmonic pollution disturbance signal C4;

[0111] (5) Flicker disturbance signal C5;

[0112] (6) Transient oscillation disturbance signal C6;

[0113] (7) Voltage gap disturbance signal C7;

[0114] (8) Voltage spike disturbance signal C8.

[0115] This embodiment strictly adheres to temporal logic through the causal convolutional structure of the temporal convolutional network, preventing the leakage of future information during online analysis. It utilizes dilated convolution to achieve exponential expansion of the receptive field with low computational cost, fully covering the temporal span of 1280 sampling points. Combined with the multi-layer feature fusion and regularization design of the residual module, the average classification accuracy reaches 98.88% in a strong electromagnetic interference environment of 30dB to 50dB, demonstrating extremely strong noise resistance and robustness, fully meeting the technical requirements of new power systems for high-precision and highly robust real-time power quality monitoring.

[0116] This embodiment, referencing the IEEE-1159 standard and considering the operational characteristics of photovoltaic grid-connected systems, generates 8000 sets of original samples using mathematical modeling. These samples include normal signals and various typical disturbances caused by inverter switching, maximum power point tracking, and sudden weather changes, with 1000 data sets for each type of disturbance. The dataset is randomly divided into training, validation, and test sets in a 7:2:1 ratio. The specific mathematical models and corresponding parameter designs for the eight types of power quality disturbance signals are shown in Table 1 below.

[0117] Table 1 Mathematical Modeling of Power Quality Disturbance Signals

[0118]

[0119] In the table, T represents the fundamental frequency period, which corresponds to 20ms for a 50Hz power grid.

[0120] Figure 2 This is a schematic diagram illustrating the principle of the discrete wavelet transform decomposition process of this invention. The input signal is alternately filtered by low-pass and high-pass filters at each level, thereby obtaining approximate and detail components layer by layer. Traditional discrete wavelet transform is often affected by shift sensitivity and boundary effects caused by downsampling when processing highly non-stationary photovoltaic grid-connected signals.

[0121] Figure 3 This is an overall flowchart of CTM-MODWT-TCN signal processing and classification provided in an embodiment of the present invention. The system first acquires the original three-phase voltage signal in real time through a 6-cycle buffer sliding window and converts it into per-unit values; then, it uses Clarke transform to map the three-phase signal to a stationary reference coordinate system, and extracts the one-dimensional CTM characteristic mode signal through feature fusion formula, achieving 66.7% data dimensionality reduction while completely preserving the unbalanced disturbance characteristics; subsequently, it performs 5-level MODWT multi-resolution decomposition on the one-dimensional CTM signal to obtain 5 high-frequency detail components (D1~D5) and 1 low-frequency approximation component (A5), which are then concatenated into a multi-channel sequence.

[0122] Figure 4This is a schematic diagram of the internal structure of the TCN core residual module in this embodiment of the invention. Each core residual block contains two cascaded dilated causal convolutional branches, utilizing a fully convolutional architecture to support large-scale parallel computation and avoid long sequence gradient vanishing. To prevent network overfitting and accelerate convergence, after each causal convolution operation, a weight normalization layer, a rectified linear unit activation layer, and a dropout layer are sequentially cascaded. The module employs a skip connection mechanism to directly add the input to the nonlinearly transformed output element-wise. When the number of input and output channels is inconsistent, convolutional layers are introduced in the skip connection branches to adjust the feature dimension of the input, ensuring smooth cross-layer transfer of lower-level features.

[0123] Figure 5 (a) shows the waveform of the original three-phase voltage signal collected by the photovoltaic grid-connected system;

[0124] Figure 5 (b) shows the α component, β component, and zero-sequence component obtained after decoupling via Clarke transform;

[0125] Figure 5 (c) shows the waveform of the one-dimensional CTM feature signal generated by feature fusion;

[0126] Figure 6 (a) shows the detail components of D1 obtained by wavelet transform decomposition;

[0127] Figure 6 (b) shows the detail components of D2 obtained by wavelet transform decomposition;

[0128] Figure 6 (c) shows the detail components of D3 obtained by wavelet transform decomposition;

[0129] Figure 6 (d) is a detail component diagram of D4 obtained by wavelet transform decomposition;

[0130] Figure 6 (e) is a detail component diagram of D5 obtained by wavelet transform decomposition;

[0131] Figure 6 In the diagram (f), the approximate component diagram of A5 obtained by wavelet transform decomposition is shown.

[0132] Figures 5-6 The waveform diagram shows the results of CTM dimensionality reduction and MODWT multi-scale decomposition processing, taking a voltage sag disturbance signal as an example, in an embodiment of the present invention. Figure 5 Images (a) and (b) show the original asymmetrical three-phase voltage waveforms and their components after Clarke coordinate transformation. Figure 5 In the middle (c), the one-dimensional CTM feature signal is aggregated after feature fusion, which fully preserves the core features of imbalance perturbation while simplifying the data dimension. Figure 6The paper demonstrates that after a one-dimensional CTM signal is decomposed into five layers of MODWT multi-resolution decomposition, five high-frequency detail components D1, D2, D3, D4, and D5, as well as one approximate component A5 representing the low-frequency trend, are extracted layer by layer.

[0133] Figure 7 Figure (a) shows the accuracy convergence curve of the TCN model of this invention during the training and validation process.

[0134] Figure 7 Figure (b) shows the loss function descent curve of the TCN model of the present invention during the training and validation process.

[0135] Figure 7 This is a graph showing the convergence curve of the TCN model during the training and validation process of the present invention. Figure 7 (a) is the accuracy convergence curve. Figure 7 (b) shows the loss function decline curve. In the model training phase, this invention sets the initial learning rate to 0.001, the batch size for a single training session to 128, the maximum number of iterations to 40, and introduces a segmented decay strategy that reduces the current learning rate to 10% every 20 iterations.

[0136] Figure 8 This is a confusion matrix diagram showing the classification test performance under different noisy conditions according to an embodiment of the present invention. Figure 8 (a)~(d) respectively demonstrate the recognition performance of the model in four independent test environments: ideal noise-free, slight noise (50dB), moderate noise (40dB), and heavy noise (30dB).

[0137] Figure 8 (a) is the classification confusion matrix of the model under an ideal, noise-free, and clean test environment;

[0138] Figure 8 (b) shows the classification confusion matrix of the model under slight noise (signal-to-noise ratio 50dB) interference.

[0139] Figure 8 (c) is the classification confusion matrix of the model under moderate noise (signal-to-noise ratio 40dB) interference environment;

[0140] Figure 8 The middle (d) diagram shows the classification confusion matrix of the model under heavy noise (signal-to-noise ratio 30dB) interference.

[0141] Secondly, this application provides a dimensionality-reduced rapid classification system for power quality disturbances in photovoltaic grid-connected systems, comprising:

[0142] The signal acquisition module is used to capture the three-phase voltage signal at the photovoltaic grid connection point in real time and perform normalization processing to obtain the three-phase voltage signal in per-unit form;

[0143] The feature extraction module is used to perform Clarke transform on the three-phase voltage signal in per-unit form to obtain a one-dimensional CTM feature waveform, and then decompose it through multi-layer maximum overlap discrete wavelet transform to generate a multi-channel time sequence matrix.

[0144] The classification module is used to load the trained temporal convolutional network model, receive the multi-channel time series matrix, and output the corresponding power quality perturbation type.

[0145] The functions of each module in the aforementioned photovoltaic grid-connected system power quality disturbance reduction and rapid classification device correspond to the steps in the aforementioned photovoltaic grid-connected system power quality disturbance reduction and rapid classification method embodiment, and their functions and implementation processes will not be described in detail here.

[0146] Thirdly, embodiments of this application also provide a readable storage medium.

[0147] This application stores a dimensionality reduction and fast classification program for power quality disturbances in a photovoltaic grid-connected system on a readable storage medium. When the dimensionality reduction and fast classification program for power quality disturbances in a photovoltaic grid-connected system is executed by a processor, it implements the steps of the dimensionality reduction and fast classification method for power quality disturbances in a photovoltaic grid-connected system as described above.

[0148] The method implemented when the dimensionality reduction and rapid classification procedure for power quality disturbances in photovoltaic grid-connected systems is executed can be referred to in various embodiments of the dimensionality reduction and rapid classification method for power quality disturbances in photovoltaic grid-connected systems in this application, and will not be repeated here.

[0149] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0150] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for rapid classification of power quality disturbances in a photovoltaic grid-connected system with dimensionality reduction, characterized in that, Includes the following steps: Collect the three-phase voltage signal of the photovoltaic grid connection point, and normalize the three-phase voltage signal to obtain the three-phase voltage signal in per-unit form; The three-phase voltage signal in per-unit form is subjected to Clarke transform to decouple it and obtain α component, β component and zero-sequence component. The α component and the β component are multiplied and then the zero-sequence component is superimposed to obtain a one-dimensional CTM characteristic waveform. The one-dimensional CTM feature waveform is decomposed by multi-level maximum overlap discrete wavelet transform to extract multiple high-frequency detail components and one low-frequency approximation component. The components are then concatenated along the channel dimension to form a multi-channel time series matrix. The multi-channel time series matrix is ​​input into the temporal convolutional network model, and the temporal dependency features are extracted through causal convolution and dilated convolution to obtain the power quality disturbance type of the corresponding photovoltaic grid-connected system.

2. The method for rapid classification of power quality disturbances in photovoltaic grid-connected systems according to claim 1, characterized in that, The normalization process for the three-phase voltage signals specifically includes: The fundamental frequency of the power grid is set to the rated power frequency, and the three-phase voltage signal is intercepted in real time through a multi-cycle buffer sliding window so that each sample contains a preset number of discrete sampling points; The amplitude of each sampling point is normalized and converted into a per-unit signal without amplitude dimension.

3. The method for rapid classification of power quality disturbances in photovoltaic grid-connected systems according to claim 2, characterized in that, The specific formula for the Clarke transform is as follows: ； In the formula, For the α component, For the β component, For zero-order components, This represents the instantaneous value of phase A voltage. This is the instantaneous value of phase B voltage. This is the instantaneous value of the C-phase voltage.

4. The method for rapid classification of power quality disturbances in photovoltaic grid-connected systems according to claim 3, characterized in that, The formula for calculating the one-dimensional CTM characteristic waveform is as follows: ； In the formula, This is a one-dimensional CTM characteristic waveform.

5. The method for rapid classification of power quality disturbances in photovoltaic grid-connected systems according to claim 1, characterized in that, In the step of performing multi-level maximum overlap discrete wavelet transform decomposition on the one-dimensional CTM feature waveform, extracting multiple high-frequency detail components and one low-frequency approximate component, and splicing each component along the channel dimension to form a multi-channel time series matrix, high-pass and low-pass filters are used alternately for filtering, and overlap is maintained between adjacent sub-bands to filter out the shift sensitivity and boundary effects caused by traditional discrete wavelet transform downsampling.

6. The method for rapid classification of power quality disturbances in photovoltaic grid-connected systems according to claim 1, characterized in that, The temporal convolutional network model contains multiple cascaded residual modules. Each residual module contains at least two cascaded dilated causal convolutions. The convolution kernel size is a preset odd number. The number of filters in each layer increases exponentially with the network depth. The dilation factor increases exponentially with the number of layers to cover the temporal span of the sampling points.

7. The method for rapid classification of power quality disturbances in photovoltaic grid-connected systems according to claim 6, characterized in that, The training process of the temporal convolutional network model includes: Set the initial learning rate, batch size for a single training session, and maximum number of iterations; During training, the learning rate is decayed in segments at preset iteration intervals to drive the model to converge to the global optimum.

8. The method for rapid classification of power quality disturbances in photovoltaic grid-connected systems according to claim 6, characterized in that, The residual module is internally connected to a weight normalization layer, a nonlinear activation layer, and a regularization layer in sequence after each dilated causal convolution operation. The input and output of the residual module are added element-wise through skip connections. When the number of input channels is inconsistent with the number of output channels, a convolutional layer is introduced in the skip connection branch to adjust the feature dimension so that the number of input and output channels matches.

9. The method for rapid classification of power quality disturbances in photovoltaic grid-connected systems according to claim 8, characterized in that, The operational logic of the causal convolution follows the following formula: ； In the formula, Let x be the output feature at the current position s, and x be the input sequence. Here, k is the kernel size, and k is the filter. This indicates that the current position s in the input signal is to the left. The value at the distance, d is the dilation rate, and i is the index of the current convolution kernel.

10. A dimension-reduced rapid classification system for power quality disturbances in a photovoltaic grid-connected system, characterized in that, include: The signal acquisition module is used to capture the three-phase voltage signal at the photovoltaic grid connection point in real time and perform normalization processing to obtain the three-phase voltage signal in per-unit form; The feature extraction module is used to perform Clarke transform on the three-phase voltage signal in per-unit form to obtain a one-dimensional CTM feature waveform, and then decompose it through multi-layer maximum overlap discrete wavelet transform to generate a multi-channel time sequence matrix. The classification module is used to load the trained temporal convolutional network model, receive the multi-channel time series matrix, and output the corresponding power quality perturbation type.

Images