Electric energy quality composite disturbance detection classification and time positioning method based on MS-TCN + +

Abstract

The invention discloses a power quality composite disturbance detection classification and time positioning method based on MS-TCN + +. According to the method, composite disturbance category codes are designed for sampling points, an electric energy quality identification model comprising a prediction generation stage and a plurality of refining stages is constructed on the basis of MS-TCN + +, and multi-time-scale disturbance signal features are extracted by using double expansion layers in the prediction generation stage. And smooth loss is added in model training to avoid sudden change of categories of sampling points in a time sequence, and finally category identification and time positioning of disturbance signals are completed through a classification result of the sampling points, so that powerful support is provided for treatment of power quality disturbance.

Description

Technical Field

[0001] The present invention belongs to the field of power systems and relates to a power quality composite disturbance detection, classification and time positioning method based on MS-TCN++. Background Art

[0002] Under the dual-carbon climate, the continuous connection of various nonlinear loads including power electronic equipment and intermittent power sources such as wind power and photovoltaics to the grid has exacerbated the occurrence and complexity of power quality disturbances (PQDs). PQDs not only occur in the form of single disturbances such as voltage interruptions, gaps, harmonics, and flicker, but also often in the form of dual or even triple compound disturbances. These frequent and increasingly complex PQDs not only affect the stable operation of the power system but can also damage electrical equipment and cause significant economic losses. Timely and accurate detection, classification, and time location of power quality disturbances will help to trace and manage these disturbances.

[0003] In recent years, PQD identification methods have been extensively studied, primarily encompassing traditional approaches based on signal processing and signal classification, as well as deep learning algorithms powered by artificial intelligence. Traditional methods are significantly influenced by the signal processing methods used to extract features, requiring manual filtering of extracted features based on empirical experience. This results in limited accuracy and generally poor performance in identifying multiple, complex PQDs. Deep learning methods, which implement automated feature extraction and classification in an end-to-end manner, have achieved promising results in PQD identification. These methods can be broadly categorized into image classification models and time series classification models. However, most of these studies only consider the presence of individual elementary perturbations within the entire waveform, without considering their temporal localization. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the defects of the above-mentioned existing technologies and provide a model that can simultaneously complete the detection, classification and time positioning of composite PQDs. By taking basic perturbations as the detection objects, hierarchical modeling of single perturbations and multiple perturbations is realized, and information such as the type, number, time interval and other information of each basic perturbation element in PQDs is obtained to obtain a more detailed description of PQDs.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] The present invention provides a method for detecting, classifying and timing complex power quality disturbances based on MS-TCN++, comprising the following steps:

[0007] S1: Generate power quality disturbance composite signal samples based on the mathematical model of power quality disturbance, and divide the samples into training set, validation set and test set in proportion;

[0008] S2: Design a composite category encoding method to represent the category of each sampling point of the power quality composite disturbance signal;

[0009] S3: Construct a power quality composite disturbance detection classification and time location model based on MS-TCN++;

[0010] S4: Use the training set to train the MS-TCN++ model. The loss function includes classification loss and smoothing loss.

[0011] S5: Post-process the point classification output of the model to obtain the final detection results, and use multiple performance indicators to evaluate the model's detection, classification, and temporal positioning performance on the test set.

[0012] The beneficial effects of the present invention are as follows: the present invention can not only complete detection and identification at the level of basic disturbance elements, but also provide the number and time interval information of each basic disturbance, which is helpful to achieve the tracing and management of power quality disturbances. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a flow chart of a method for detecting, classifying and timing complex power quality disturbances based on MS-TCN++ according to an embodiment of the present application.

[0014] Figure 2 This is the overall structure of the MS-TCN++ model in the embodiment of this application.

[0015] Figure 3 This is a structural diagram of a single dilated convolutional network according to an embodiment of the present application.

[0016] Figure 4 This is a structural diagram of the double-expanded convolutional network of an embodiment of this application.

[0017] Figure 5 This is a comparison chart of the time positioning errors of the amplitude disturbance and oscillation disturbance of the embodiment model of this application at different noise levels.

[0018] Figure 6 This is a comparison chart of the time positioning error of transient basic disturbances under different noise levels of the embodiment model of this application. DETAILED DESCRIPTION

[0019] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0020] like Figure 1As shown, the embodiment of the present application provides a method for detecting, classifying and timing complex power quality disturbances based on MS-TCN++, comprising the following steps:

[0021] S1: Generate power quality disturbance composite signal samples based on the mathematical model of power quality disturbance, and divide the samples into training set, validation set and test set in proportion.

[0022] Furthermore, there are 64 samples of power quality disturbance composite signals, including 11 single disturbance signals, 28 double disturbance signals, and 25 triple disturbance signals, among which the normal signal is regarded as a special single disturbance signal.

[0023] Table 1 shows the mathematical models of 10 basic power quality disturbance elements in four categories.

[0024] Table 1 Basic power quality disturbance classification table

[0025]

[0026] The composite perturbation is generated according to formula (1). To ensure the quality of samples, only one perturbation is considered for each category.

[0027]

[0028] Where sin(t) is a standard sinusoidal signal; sag(t), swl(t), and itr(t) are the disturbance factors of voltage sag, voltage swell, and voltage interruption, respectively; imp(t), osc(t), and nth(t) are the disturbance factors of transient impulse, transient oscillation, and voltage notch, respectively; flk(t) and hmy(t) are the disturbance factors of flicker and harmonics, respectively; and noise(t) is the noise signal.

[0029] The samples were divided into training, validation, and test sets in a ratio of 4:1:1. The training and validation sets contained equal proportions of samples with no noise, a 50dB SNR, a 40dB SNR, and a 30dB SNR for each disturbance signal. Noise signals were added to the test set based on test needs.

[0030] S2: Design a composite category encoding method to represent the category of each sampling point of the power quality composite disturbance signal.

[0031] The embodiment of the present application proposes a hybrid encoding method for the sampling points of the PQDs sequence, which encodes each sampling point of the disturbance signal sequence separately, and the category label of the sampling point is represented by the category combination of the basic disturbance in which the sampling point is located. Specifically:

[0032] First, multi-label classification encoding is used to decouple the sampling point classification task into four subtasks, corresponding to the identification of four basic disturbance categories: transient disturbances, amplitude disturbances, flicker disturbances, and harmonic disturbances. Because disturbances of the same category do not occur simultaneously, multi-class encoding is employed within transient and amplitude disturbances. The encoding vectors corresponding to the four subtasks form a 12-length encoding vector, representing the composite category of the sampling point. Transient disturbances are encoded using a one-hot vector of length 6, with the six elements corresponding to normal, spike, notch, oscillation, positive pulse, and negative pulse, respectively. Similarly, amplitude disturbances are represented using a one-hot vector of length 4, with the four elements corresponding to normal, swell, sag, and interruption, respectively. Flicker and harmonic disturbances are both labeled using a 1-0 encoding of length 1 to indicate presence or absence. Furthermore, the "normal" category in the one-hot vectors for transient and amplitude disturbances indicates the absence of any of the basic disturbances from the corresponding category at the sampling point.

[0033] S3: Construct a power quality composite disturbance detection, classification, and time location model based on MS-TCN++, which includes a single-dilated time convolutional network, a double-dilated time convolutional network, and a probabilistic activation function.

[0034] like Figure 3 As shown, the convolution kernel size of the first convolution layer of the single dilated time convolutional network is 1×1, which is used to map the dimensionality of the input sequence to the dimensionality of the hidden layer features; the final 1×1 convolution layer is used to map the dimensionality of the hidden layer features to the dimensionality of the output layer features. In this embodiment of the application, the dimension of the output layer features is 12, corresponding to the encoding vector of length 12 described in step S2.

[0035] In the middle of the network is a stack of L single dilated convolution residual blocks. Each single dilated convolution residual block is composed of a one-dimensional dilated convolution layer, a ReLU activation layer, a 1×1 convolution layer, and a Dropout layer in series. The residual connection structure is introduced to alleviate the problem of gradient disappearance.

[0036] The convolution kernel size of all one-dimensional dilated convolutions is 1×3, and the dilation coefficient of the lth single dilated convolution layer is 2 l ; The ReLU activation layer increases the nonlinear fitting ability of the neural network; the 1×1 convolution layer in the residual block is used to model the feature channel; Dropout is a regularization technique that only works during model training.

[0037] During the model training process, the Dropout layer makes each neuron have a certain probability of failure, and the combination of neurons retained in each forward propagation is different. This can greatly reduce the dependency between features and enable the model to learn more robust features, thereby achieving a regularization effect. The probability of neuron loss in the embodiment of the present application is taken as 0.5.

[0038] The double dilated time convolution network adds a new dilated convolution layer based on the single dilated time convolution network. Its structure is as follows: Figure 4 Specifically, the double-dilated convolution residual block contains two dilated convolution layers, one with an increasing dilation coefficient and the other with a decreasing dilation coefficient. The convolution results of the two layers are fused by splicing, extracting features at multiple time scales. The rest of the structure is the same as the single-dilated temporal convolution network.

[0039] When the neural network outputs the composite category probability for each sampling point, different output activation functions are needed to map the output vector into a probability form. The first six elements and the seventh to tenth elements of the output vector are activated by Softmax to output the category probability predictions of transient disturbances and amplitude disturbances, respectively. The last two elements use the Sigmoid activation function to obtain the probability predictions of flicker disturbances and harmonic disturbances, respectively. The expressions of the Softmax and Sigmoid activation functions are shown in Equations (2) and (3), respectively:

[0040]

[0041] Where x i is the i-th element of the output vector belonging to transient disturbance or amplitude disturbance, N c is the class label length, which is 6 and 4 for transient disturbances and amplitude disturbances respectively. The role of the Softmax function is to convert each element in the input vector into its form as a probability value, while Sigmoid converts a single element into a probability value.

[0042] like Figure 2 As shown, the embodiment of the present application proposes an MS-TCN++ network, which includes a prediction generation stage and multiple refinement stages.

[0043] The prediction generation stage takes the one-dimensional sequence of PQDs as input. After passing through the double-dilated time convolutional network, the output vector is mapped into the form of probability using the output activation function of Equations (2) and (3) to obtain the initial category prediction results of each sampling point in the PQDs sequence, where the probability activation operation is represented by Map in the figure.

[0044] Each subsequent refinement stage takes the predicted probability of the perturbation category of each sampling point in the PQDs sequence output in the previous stage as input, and uses a single dilated time convolutional network to continuously optimize the predicted probability. The number of refinement stages is represented by N in the figure. r Indicates that the input-output relationship of each stage in the MS-TCN++ model is shown in formula (4):

[0045]

[0046] Where s is the number of stages; F represents the temporal convolutional network described in Section 3.1. Only when s = 1 is F a double-dilated temporal convolutional network, and in the remaining stages F is a single-dilated temporal convolutional network; M represents the probabilistic activation operation; z 1:T is the PQDs sequence; T is the length of the input time series; Z s is the output result of the network in the sth stage; Y s is the category probability prediction result of the s-th stage sequence.

[0047] S4: Use the training set to train the MS-TCN++ model. The loss function includes classification loss and smoothing loss.

[0048] Each stage of the temporal convolutional network has a corresponding loss value, and the total loss of the model is L M is the sum of the loss values ​​of each stage, and the loss value of each stage includes the classification loss L cls and smoothing loss L smooth Two parts, expressed as formula (5):

[0049]

[0050] Where, L s is the loss value of the model at the sth stage; S is the total number of stages of the model; λ is the weight coefficient of the smooth loss.

[0051] The classification loss is calculated for the sub-labels corresponding to the four major types of disturbances. The classification loss of transient disturbances and amplitude disturbances uses multi-classification cross entropy loss, while the classification loss of flicker and harmonic disturbances uses binary cross loss, as shown in Equations (6) and (7):

[0052]

[0053] Where, is the sum of the classification losses corresponding to the i-th category of disturbance at each sampling point in the sequence; i=1,2 is the predicted probability given by the model at time step t for the true disturbance category c in the i-th category of disturbance; i=3,4 is the model's predicted probability value for flicker or harmonic disturbance at time step t; i=3, 4 are the corresponding label values.

[0054] Smoothing loss penalizes the difference in prediction probabilities of adjacent sampling points, reduces the jump in the prediction probabilities of adjacent sampling points, and thus reduces the occurrence of misclassification. Smoothing loss uses the mean square error of truncation on the logarithmic probability, and its expression is as follows:

[0055]

[0056] △ t,c =|lny t,c -lny t-1,c | (10)

[0057] Among them, the smoothing loss is calculated in the same way for the 10 basic disturbances and the two normal categories of amplitude disturbance and transient disturbance. Therefore, in Equations (8), (9) and (10), the predicted probability y of the sampling point category c at time step t is t,c No superscript is used to distinguish them. C is the total number of categories, 12; τ is the cutoff threshold of the smoothing loss.

[0058] S5: Post-process the point classification output of the model to obtain the final detection results, and use multiple performance indicators on the test set to evaluate the detection, classification, and positioning performance of the model.

[0059] The MS-TCN++ model ultimately reports the probability of each basic perturbation in each of the four categories for each sampling point in the PQDs sequence. To determine the type and time interval of each basic perturbation in the PQDs signal, the point classification results must be processed as follows.

[0060] For power quality composite disturbance signals containing transient and amplitude disturbances, since the disturbances within a broad category do not occur at the same time, the most probable category within each category is used as the disturbance category label for that sampling point. Based on the change in the sampling point category label over time, the entire PQDs segment is divided into multiple time segments, with the start and end times of each time segment corresponding to the moment when the sampling point category changes.

[0061] For power quality composite disturbance signals containing flicker and harmonic disturbances, since these are long-term steady-state disturbances, we do not investigate their temporal location; instead, we only detect their occurrence. The flicker or harmonic disturbance confidence level for all sampling points in the signal is averaged to represent the probability of occurrence of the corresponding disturbance for the entire signal segment. A probability confidence level greater than 0.5 indicates the presence of the disturbance.

[0062] The detection and classification performance of the model is evaluated using precision, recall, and F1-Score. The expressions of the three are as follows:

[0063]

[0064] Where TP, FP, and FN represent the number of correct identifications, false identifications, and missed identifications, respectively. Precision is the proportion of true targets among all predictions given by the model, measuring the model's ability to find only relevant targets. Recall indicates the maximum number of true targets covered by the model's predictions, measuring the model's ability to find all relevant targets. The F1 score is the harmonic mean of the two and comprehensively evaluates the model's detection performance.

[0065] The mean absolute error η of the start / end time of the correctly identified samples TP of each type of basic perturbation is calculated as follows: b and η e :

[0066]

[0067] Where TP is the number of correct recognitions, f s is the sampling frequency, and Respectively represent the starting point and end point coordinate prediction values ​​of each TP, and The real values ​​of the coordinates of the start and end points respectively.

[0068] To verify the effectiveness of the MS-TCN++-based power quality composite disturbance detection, classification, and time location method described in this application, this application also provides an application example. The 64 power quality disturbance signals shown in Table 2 were generated, including 11 single disturbance signals, 28 double disturbance signals, and 25 triple disturbance signals. Normal signals were considered a special single disturbance signal.

[0069] Table 2 PQDs signal types used in the sample set

[0070]

[0071] In the dataset, the fundamental frequency was set to 50 Hz, the sampling frequency was set to 6.4 kHz, and each PQD sequence was sampled for 0.28 seconds, totaling 1792 sampling points. Each perturbation type contained 2400 samples. The dataset was then divided into training, validation, and test sets in a 4:1:1 ratio. The training and validation sets had equal proportions of noise-free, 50dB SNR, 40dB SNR, and 30dB SNR samples for each perturbation signal. Noise signals were added to the test set based on subsequent experiments.

[0072] After comparison and debugging, the total number of stages of the model is 4, including 1 prediction generation stage and 3 refinement stages. The number of double-dilated convolution residual blocks in the prediction generation stage is 11, and the number of dilated convolution residual blocks in the refinement stage is 10. The feature dimension of the hidden layer is uniformly set to 64, and the smoothing loss weight coefficient λ and truncation value τ are taken as 0.05 and 4 respectively.

[0073] In the training, validation, and testing of the model, the batch size is 64, the optimizer is Adam, and the initial learning rate is 5×10 -4 The learning rate decays to 0.9 of the previous generation after each training epoch. An early stopping mechanism is used during training to prevent overfitting. Training is stopped when the validation loss of five consecutive generations of models exceeds the historical minimum. The model with the lowest validation loss is selected as the final model. A total of 44 epochs were trained, so the model from the 39th epoch is selected as the final model.

[0074] Noise with different signal-to-noise ratios was added to the test set, and the F1 score indicator was tested. The results are shown in Table 3.

[0075] Table 3 F1 scores at different noise levels

[0076]

[0077]

[0078] It can be seen that under the above signal-to-noise ratio environment, the comprehensive F1 score index of the model of this application is above 99.2%, and there are fewer false detections and missed detections.

[0079] The mean absolute error η of the start / end times is calculated using the correctly identified samples of each type of basic perturbation b and η e As an evaluation indicator of the model's time localization task at the basic perturbation level.

[0080] The start / end time positioning error statistics of the correctly identified samples are performed, and the η of the eight basic perturbations with clear start and end times in different signal-to-noise ratio test sets is b ,η e like Figure 5 and Figure 6 shown.

[0081] Figure 5 and Figure 6 The figure shows the mean absolute error (MAE) for the start and end times of eight basic perturbations, with the vertical axis measured in sampling points (pts). At a sampling frequency of 6.4kHz, an error of one sampling point corresponds to 0.156ms. The figure shows that, under varying noise levels, the model's MAE for locating the start and end times of amplitude perturbations is consistently less than one sampling point.

[0082] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not limitations on the implementation methods of the present invention. For ordinary technicians in this field, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the implementation methods here. All obvious changes or modifications derived from the technical solution of the present invention are still within the scope of protection of the present invention.

Claims

1. A power quality composite disturbance detection, classification and time location method based on MS-TCN++, characterized by: The following steps are involved: Step 1: Generate power quality disturbance composite signal samples based on the mathematical model of power quality disturbance and divide the samples into training set, validation set and test set; Step 2: Design a composite category encoding method to represent the category of each sampling point of the power quality composite disturbance signal; Step 3: Construct a power quality composite disturbance detection classification and time location model based on MS-TCN++; Step 4: Train the MS-TCN++ model using the training set; Step 5: Post-process the point classification output of the model to obtain the final detection results.

2. The power quality composite disturbance detection, classification and time positioning method based on MS-TCN++ according to claim 1 is characterized in that: The power quality disturbance composite signal sample includes a single disturbance signal, a double disturbance signal and a triple disturbance signal; wherein a normal signal is regarded as a special single disturbance signal.

3. The power quality composite disturbance detection, classification and time positioning method based on MS-TCN++ according to claim 1 or 2, characterized in that: In step 1, the samples are divided into a training set, a validation set, and a test set in a ratio of 4:1:1; wherein each disturbance sample in the training set and the validation set contains four noise levels: no noise, signal-to-noise ratio of 50dB, signal-to-noise ratio of 40dB, and signal-to-noise ratio of 30dB, and the proportions are the same.

4. The power quality composite disturbance detection, classification and time positioning method based on MS-TCN++ according to claim 1 is characterized in that: The step 2 is specifically as follows: Multi-label classification coding is adopted to decouple the sampling point classification task into four subtasks, corresponding to the identification tasks of four basic disturbances: transient disturbance, amplitude disturbance, flicker disturbance and harmonic disturbance. Multi-classification coding is adopted within transient disturbance and amplitude disturbance. The coding vectors corresponding to the four subtasks are combined into a total coding vector to represent the composite category of the sampling point.

5. The power quality composite disturbance detection, classification and time positioning method based on MS-TCN++ according to claim 1 is characterized in that: The MS-TCN++ model includes a single dilated temporal convolutional network, a double dilated temporal convolutional network, and a probabilistic activation function; The double-dilated temporal convolutional network is used to obtain the initial category prediction results of each sampling point in the PQDs sequence; The single dilated temporal convolutional network is used to optimize the prediction probability; The probability activation function is used to obtain the probability of each basic disturbance type at the sampling point.

6. The power quality composite disturbance detection, classification and time positioning method based on MS-TCN++ according to claim 5 is characterized in that: The single dilated time convolutional network has a 1×1 convolutional layer at the beginning and end, respectively. The first convolutional layer is used to map the dimension of the input sequence to the dimension of the hidden layer features, and the second convolutional layer is used to map the dimension of the hidden layer features to the dimension of the output layer features. The middle of the network is a stack of L single dilated convolution residual blocks, where L is a set value. Each single dilated convolution residual block is composed of a one-dimensional dilated convolution layer, a ReLU activation layer, a 1×1 convolution layer, and a Dropout layer connected in series.

7. The power quality composite disturbance detection, classification and time positioning method based on MS-TCN++ according to claim 6 is characterized in that: The double dilated temporal convolutional network adds a new dilated convolutional layer on the basis of the single dilated temporal convolutional network; There are two dilated convolution layers in the double dilated convolution residual block. The dilation coefficient of one layer increases with the number of layers, while the dilation coefficient of the other layer decreases with the number of layers. The convolution results of the two layers are fused by splicing. The rest of the structure is the same as the single dilated temporal convolutional network.

8. The method for detecting, classifying and timing-locating power quality composite disturbances based on MS-TCN++ according to claim 5, 6 or 7, characterized in that: The working method of the MS-TCN++ model is: First, in the prediction generation stage, the one-dimensional sequence of PQDs is used as input. After passing through the double-dilated temporal convolutional network, the output activation function is used to map the output vector into the form of probability, and the initial category prediction results of each sampling point in the PQDs sequence are obtained. Each subsequent refinement stage takes the predicted probability of the perturbation category of each sampling point in the PQDs sequence output in the previous stage as input, and uses a single dilated time convolutional network to continuously optimize the predicted probability.

9. The power quality composite disturbance detection, classification and time positioning method based on MS-TCN++ according to claim 1, characterized in that: The step 4 is specifically as follows: The MS-TCN++ model is trained using the training set. The total loss of the model is the sum of the loss values ​​corresponding to the temporal convolutional network at each stage. The loss value of each stage includes classification loss and smoothing loss: Among them, L M is the total loss of the model, L s is the loss value of the s-th stage model, S is the total number of stages of the model, L cls is the classification loss, L smooth is the smoothing loss, and λ is the weight coefficient of the smoothing loss.

10. The power quality composite disturbance detection, classification and time positioning method based on MS-TCN++ according to claim 1, characterized in that: The step 5 further comprises: Use performance indicators on the test set to evaluate the detection, classification, and temporal localization performance of the model; The performance indicators include evaluation indicators for detection and classification tasks and evaluation indicators for time positioning tasks. Precision, recall and F1 score are used as evaluation indicators for the model's detection and classification tasks at the basic perturbation level. The mean absolute error of the start / end time of correctly identified samples of each type of basic perturbation is used as an evaluation indicator for the model's time positioning tasks at the basic perturbation level.

Images