Harmonic source modeling method based on time-frequency feature extraction

Abstract

The invention provides a harmonic source modeling method based on time-frequency feature extraction. The method comprises the following steps: firstly, determining harmonic principal component frequency of a harmonic source; extracting harmonic voltage amplitude and phase angle corresponding to harmonic principal component frequency of the harmonic source to obtain a harmonic source frequency domain feature set; reconstructing the frequency domain feature set through IDFT to form a harmonic source frequency domain feature component matrix, and splicing the harmonic source frequency domain feature component matrix with an original voltage sequence acquired by a synchronous phasor measurement device to obtain a time-frequency feature component matrix; converting the component matrix into a grey-scale map by using a matrix normalization method, and converting the grey-scale map into a three-channel time-frequency characteristic component color map by using a pseudo-color coding technology; and finally, taking the acquired harmonic source voltage data and the time-frequency characteristic component color map as the input of the model, taking the acquired corresponding harmonic source current data as the output of the model, and continuously training the network model through the actually measured data until the accurate mapping of the relation between the harmonic voltage and the harmonic current of the harmonic source is realized. And obtaining a harmonic source data driving model.

Description

Technical Field

[0001] The present invention relates to the technical fields of power big data, harmonic source modeling, and in particular to a harmonic source modeling method based on time-frequency feature extraction. Background Art

[0002] With the transformation of energy structure and the rapid improvement of national economic industrialization, the number of nonlinear loads such as arc furnaces, electric locomotives and power electronic equipment has increased sharply, and the resulting harmonic pollution problem has become increasingly serious. Accurately grasping the harmonic characteristics of harmonic source loads is helpful for harmonic pollution prevention and control and improving the quality of power. Therefore, it is of great significance to construct an accurate and reasonable harmonic source model.

[0003] At present, harmonic source modeling can be divided into two categories: frequency domain modeling and time domain modeling according to the form of the established harmonic voltage and harmonic current expressions. The frequency domain harmonic source model can output the vector value of each harmonic emitted by the harmonic source, which can reflect the long-term steady-state frequency domain characteristics of the harmonic source. However, this type of model usually needs to update the model parameters when the operating state of the harmonic source changes. Therefore, it is often used for harmonic sources with relatively stable working conditions, and it is difficult to apply to situations where the working conditions of the harmonic source change frequently or the harmonic source has strong time-varying properties. Time domain modeling reflects the time-varying characteristics of the harmonic source by introducing timing information. The resulting harmonic source voltage and current time domain waveforms contain the instantaneous change characteristics of the harmonic source. It is widely used in establishing accurate harmonic source models, but it is often difficult to ensure modeling accuracy and generalization for complex harmonic sources with unknown specific topology and unclear mechanism characteristics.

[0004] In the context of a wide variety of harmonic sources with different properties, the harmonic characteristics of harmonic sources are greatly affected by the mechanism of generating harmonics. If only the dynamic time-varying characteristics of the harmonic source are concerned, the harmonic source model constructed will be too complex, lack generalization, and easily overfit; if only the frequency domain steady-state characteristics of the harmonic source are concerned, the harmonic characteristics of the harmonic source running in a non-steady state cannot be accurately reflected, and it is difficult to characterize the time-varying nature of the harmonic source. In order to more comprehensively reflect the harmonic characteristics of the harmonic source and improve the accuracy of the harmonic source model, the short-term time-varying characteristics and long-term steady-state frequency domain characteristics of the harmonic source should be extracted simultaneously for analysis and modeling. At present, some scholars have added constraints on the frequency domain to the minimized objective error function of the RBF neural network, and improved the model's anti-interference ability to time-varying noise by introducing frequency domain features. However, the frequency domain error function in this method is complex and difficult to solve, and the corresponding parameters of the time domain and frequency domain error functions are not easy to determine in the update of neuron weights. It can be seen that how to take into account the steady-state frequency domain characteristics and dynamic time-varying characteristics of the harmonic source to improve the accuracy and generalization of the current harmonic source modeling is a major problem in the current harmonic source modeling. Summary of the invention

[0005] In view of the limitations of the existing technology, considering that with the widespread application of synchronized phasor measurement devices, a large amount of accurate harmonic source measured data has been continuously accumulated, and the data-driven harmonic source time domain modeling method has made certain progress. This type of method starts from the external characteristics of the harmonic source, without knowing the specific topological structure, and only requires the measurement data of the harmonic source port to establish the harmonic source time domain model. Among them, deep learning can achieve better results in the time domain modeling of harmonic sources with strong nonlinearity and complex and unclear internal mechanisms by relying on the powerful nonlinear mapping ability of neural networks. In view of the fact that the existing synchronized phasor measurement devices can already provide abundant high-precision harmonic measurement data, the present invention proposes a harmonic source data-driven modeling method based on time-frequency feature extraction.

[0006] This method is based on the time domain analysis method and adopts a heterogeneous feature input combination that reflects the dynamic time-varying characteristics of the harmonic source and the steady-state characteristics of the frequency domain. It introduces a harmonic source voltage-current mapping model that combines multiple convolutional neural networks (MCNN) and bidirectional long short term memory networks (BiLSTM).

[0007] The design points of the scheme include:

[0008] (1) A time-frequency feature extraction method is proposed. By analyzing the principal components of the harmonics of the harmonic source, the key frequency domain features that affect the harmonic characteristics of the harmonic source can be accurately extracted. By constructing a color map of the time-frequency feature components, the time-frequency features are enhanced and fused.

[0009] (2) An MCNN model is proposed, in which the time domain feature extraction module and the frequency domain feature extraction module jointly extract features. The two-way parallel structure can simultaneously extract different feature information of multi-channel and multi-dimensional inputs.

[0010] (3) An MCNN-BiLSTM model is proposed. The MCNN and BiLSTM are connected in series. The time-frequency domain fusion features extracted from the MCNN model are input into the BiLSTM network to obtain the characteristic information of the harmonic voltage in the entire time domain.

[0011] The specific technical solutions are as follows:

[0012] A harmonic source modeling method based on time-frequency feature extraction:

[0013] Firstly, the voltage and current data of the harmonic source collected by the synchronous phasor measurement device are normalized and analyzed by FFT; the content rate of each harmonic component is calculated, and the frequency of the main harmonic component of the harmonic source is determined by the set threshold judgment rule;

[0014] Then, the harmonic voltage amplitude and phase angle corresponding to the harmonic main component frequency of the harmonic source are extracted to obtain the frequency domain feature set of the harmonic source; the frequency domain feature set is reconstructed by IDFT to form a frequency domain feature component matrix of the harmonic source, and it is spliced ​​with the original voltage sequence collected by the synchronous phasor measurement device to obtain a time-frequency feature component matrix; the component matrix is ​​converted into a grayscale image using a matrix normalization method, and is converted into a three-channel time-frequency feature component color image using a pseudo-color coding technique;

[0015] Finally, the harmonic source voltage data collected by the synchronized phasor measurement device and the time-frequency characteristic component color map are used as the input of the MCNN-BiLSTM model, and the corresponding harmonic source current data collected by the synchronized phasor measurement device are used as the output of the MCNN-BiLSTM model. The network model is continuously trained through measured data until the accurate mapping of the relationship between the harmonic voltage and harmonic current of the harmonic source is achieved, thereby obtaining a harmonic source data-driven model.

[0016] The MCNN-BiLSTM model is a harmonic source voltage and current mapping model composed of a multiple convolutional neural network MCNN and a bidirectional long short-term memory network BiLSTM.

[0017] Furthermore, in the MCNN-BiLSTM model, the MCNN model extracts features jointly by a time domain feature extraction module and a frequency domain feature extraction module, and the two-way parallel structure realizes the simultaneous extraction of different feature information of multi-channel and multi-dimensional inputs; by serially connecting MCNN and BiLSTM, the MCNN model is extracted into time-frequency domain fusion features and input into the BiLSTM network to obtain the feature information of harmonic voltage in the entire time domain.

[0018] Furthermore, the specific steps of harmonic principal component analysis of harmonic sources are as follows:

[0019] First, the raw data obtained by the synchronized phasor measurement device is normalized to unify the scale of the raw measurement data of the harmonic source:

[0020]

[0021] Where: is the normalized voltage and current sequence, U and I are the original voltage and current sequences measured by the synchronous phasor measurement device, and U max ,I max The extreme values ​​of the original voltage and current sequence measured by the synchronous phasor measurement device;

[0022] Secondly, the normalized voltage and current sequence is subjected to fast Fourier transform to obtain the corresponding k-th order harmonic component:

[0023]

[0024] Where: U k , is the k-th order harmonic voltage amplitude and phase angle; I k , is the amplitude and phase angle of the k-th order harmonic current;

[0025] Then, calculate the harmonic voltage content rate HRU k Harmonic current content HRI k :

[0026]

[0027] Where: U1 and I1 are the fundamental voltage and fundamental current amplitudes respectively;

[0028] Finally, all HRUs are selected according to the following formula k and HRI k The harmonic frequencies that account for more than 1% are taken as the harmonic principal component frequencies, forming the harmonic principal component frequency set γ of the harmonic source;

[0029]

[0030] Furthermore, according to the harmonic main component order set, the corresponding k-th harmonic voltage amplitude and phase angle are extracted to obtain the harmonic source frequency domain feature set: The steps to convert F into a time-frequency feature component color map are as follows:

[0031] First, the inverse Fourier transform is used to reconstruct the harmonic voltage components:

[0032]

[0033] Where: U k (t) represents the kth harmonic voltage characteristic component, N is the number of sampling points;

[0034] Secondly, the harmonic voltage characteristic components {U1(t),U3(t),…,U k (t)} is concatenated with the original voltage sequence U(t) collected by the synchronous phasor measurement device to obtain the time-frequency characteristic component matrix {U1(t),U3(t),…,U k (t),U(t)};

[0035] Then, the time-frequency feature component matrix is ​​processed using the following formula, so that its data are all normalized to the range of (0, 255), and then it is matched with the grayscale value to obtain the grayscale image of the time-frequency feature;

[0036]

[0037] Where: G(m,n) represents the value of the gray matrix in the mth row and nth column; x i represents the value of the i-th data point; x min With x max Respectively represent the minimum and maximum values ​​in the time-frequency feature component matrix;

[0038] Finally, the grayscale image is converted into a three-channel RGB color image using pseudo color coding to obtain a color map of the time-frequency feature components.

[0039] Furthermore, the time domain feature extraction module adopts a double-layer 1D convolutional network structure, each layer of the 1D convolutional network includes a 1D convolutional layer, a 1D regularization layer and a 1D pooling layer; the harmonic source voltage data X1 collected by the synchronized phasor measurement device is used as input, and passes through two layers of 1D convolutional networks in sequence to finally obtain the feature vector Y 1DCNN :

[0040] Y 1DCNN =F 1DCNN (X1) (19)

[0041] Where: F 1DCNN Represents the time domain feature extraction module.

[0042] Furthermore, the frequency domain feature learning module adopts a double-layer 2D convolutional network, each layer of the 2D convolutional network includes a 2D convolutional layer, a 2D regularization layer and a 2D pooling layer; the time-frequency feature component color map X2 is used as input, and passes through two layers of 2D convolutional networks in sequence, and finally obtains the feature vector Y 2DCNN :

[0043] Y 2DCNN =F 2DCNN (X2) (20)

[0044] Where: F 2DCNN Represents the frequency domain feature extraction module.

[0045] Furthermore, the feature vectors output by the time domain feature extraction module and the frequency feature extraction module are flattened and converted into two feature vectors of the same size through the fully connected layer; the two feature vectors are fused by splicing to obtain the fused feature Y cat , as the local deep representation of the time-frequency features of the harmonic source characteristics, the splicing operation is expressed as follows:

[0046]

[0047] Where: Y cat It represents the fusion features after splicing, flatten is the flattening function, and Fc is the fully connected layer function.

[0048] Furthermore, the input of the BiLSTM network is the fusion feature Y cat , the output of the model is the harmonic current X out , specifically:

[0049] X out =F BiLSTM (Y cat ) (twenty two)

[0050] Where: F BiLSTM Represents a BiLSTM network.

[0051] And, an electronic device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the computer program comprises a harmonic source data-driven model obtained by the above-mentioned harmonic source modeling method based on time-frequency feature extraction.

[0052] A non-transitory computer-readable storage medium stores a computer program thereon, wherein the computer program includes a harmonic source data-driven model obtained by the above-mentioned harmonic source modeling method based on time-frequency feature extraction.

[0053] Compared with the prior art, the present invention and its preferred solution can give full play to the advantages of high-quality harmonic acquisition data of synchronized phasor measurement devices, realize data-driven accurate modeling of harmonic source models, and provide support for harmonic pollution prevention and control and improvement of power quality levels. The time-frequency feature extraction method proposed in the present invention can accurately extract the harmonic main component frequency that affects the harmonic characteristics of the harmonic source, and use IDFT and pseudo-color coding technology to perform feature dimension upgrade on discrete frequency domain features and fuse time domain features. The resulting time-frequency feature color map has richer spatiotemporal feature information and stronger feature expression than the 1-dimensional time domain waveform. The proposed MCNN model uses convolution blocks of different scales and convolution kernels of different sizes to fully explore the correlation between short-term local features of the harmonic source in the time domain and harmonic components of each frequency in the frequency domain. Compared with the single-structure CNN model, MCNN can extract richer and more comprehensive harmonic source feature information. The MCNN-BiLSTM model has better model accuracy in both single harmonic source modeling and complex harmonic source network modeling, and can more comprehensively reflect the harmonic characteristics of the harmonic source. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0055] Figure 1 This is a schematic diagram of the model building process and network structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0056] Hereinafter, the specific embodiments of the present application will be described in detail with reference to the accompanying drawings. According to these detailed descriptions, those skilled in the art can clearly understand the present application and can implement the present application. Without violating the principles of the present application, the features in the various embodiments can be combined to obtain new implementations, or certain features in certain embodiments can be replaced to obtain other preferred implementations.

[0057] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0058] In order to make the features and advantages of this patent more obvious and easy to understand, the following embodiments are specifically cited and described in detail as follows: 1 Harmonic source harmonic principal component analysis

[0059] In order to more accurately extract the frequency domain characteristics of the harmonic source, it is necessary to determine the main frequency components of the harmonics emitted by the harmonic source. The specific steps are as follows:

[0060] First, the original data obtained by the synchronized phasor measurement device is normalized to unify the scale of the original measurement data of the harmonic source. The specific formula is as follows:

[0061]

[0062] Where: is the normalized voltage and current sequence, U and I are the original voltage and current sequences measured by the synchronous phasor measurement device, and U max ,I max It is the extreme value of the original voltage and current sequence measured by the synchronous phasor measurement device.

[0063] Secondly, the normalized voltage and current sequence is subjected to fast Fourier transform (FFT) to obtain the corresponding k-th order harmonic component:

[0064]

[0065] Where: U k , is the k-th order harmonic voltage amplitude and phase angle; I k , is the amplitude and phase angle of the k-th order harmonic current.

[0066] Then, calculate the harmonic voltage content rate HRU kHarmonic current content HRI k :

[0067]

[0068] Where: U1 and I1 are the fundamental voltage and fundamental current amplitudes respectively.

[0069] Finally, all HRUs are selected according to the following formula k and HRI k The harmonic frequencies with a proportion greater than 1% are taken as the harmonic principal component frequencies, forming the harmonic source harmonic principal component frequency set γ.

[0070]

[0071] 2. Time-frequency characteristic component color map

[0072] According to the harmonic principal component order set obtained in Section 1, the corresponding k-th harmonic voltage amplitude and phase angle are extracted to obtain the harmonic source frequency domain feature set The steps to convert F into a time-frequency feature component color map are as follows:

[0073] First, the inverse discrete Fourier transform (IDFT) is used to reconstruct the harmonic voltage components. The specific formula is as follows:

[0074]

[0075] Where: U k (t) represents the kth harmonic voltage characteristic component, and N is the number of sampling points.

[0076] Secondly, the harmonic voltage characteristic components {U1(t),U3(t),…,U k (t)} is concatenated with the original voltage sequence U(t) collected by the synchronous phasor measurement device to obtain the time-frequency characteristic component matrix {U1(t),U3(t),…,U k (t),U(t)}.

[0077] Then, the time-frequency feature component matrix is ​​processed using the following formula, so that its data are all normalized to the range of (0, 255), and then it is matched with the grayscale value encoded inside the computer to obtain a grayscale image of the time-frequency feature.

[0078]

[0079] Where: G(m,n) represents the value of the gray matrix in the mth row and nth column; x i represents the value of the i-th data point; x min With xmax Respectively represent the minimum and maximum values ​​in the time-frequency feature component matrix.

[0080] Finally, the grayscale image is converted into a three-channel RGB color image using pseudo color coding technology to obtain a time-frequency feature component color map.

[0081] 3MCNN-BiLSTM network model

[0082] 3.1 MCNN Network

[0083] Convolutional neural network is a deep learning algorithm. The convolution layer it contains can extract local features of input data through convolution operation, and has good performance in processing multi-dimensional nonlinear data. Since the setting of the convolution kernel greatly affects the quality of feature extraction, given that the time domain features of the harmonic source present complex time-varying characteristics while the frequency domain features are relatively stable, in order to improve the effectiveness of feature extraction, the present invention proposes a time domain feature extraction module and a frequency domain feature extraction module.

[0084] (1) Time domain feature extraction module

[0085] The time domain feature learning module adopts a two-layer 1D convolutional network (1DCNN) structure. Each layer of 1DCNN contains a 1D convolution layer, a 1D regularization layer, and a 1D pooling layer. The harmonic source voltage data X1 collected by the synchronized phasor measurement device is used as input and passes through two layers of 1DCNN in sequence. Finally, the feature vector Y 1DCNN , specifically:

[0086] Y 1DCNN =F 1DCNN (X1) (29)

[0087] Where: F 1DCNN Represents the time domain feature extraction module.

[0088] (2) Frequency domain feature extraction module

[0089] The frequency domain feature learning module uses a two-layer 2D convolutional network (2DCNN). Each layer of 2DCNN contains a 2D convolution layer, a 2D regularization layer, and a 2D pooling layer. The time-frequency feature component color map X2 obtained in Section 2 is used as input and passes through two layers of 2DCNN in sequence. Finally, the feature vector Y 2DCNN , specifically:

[0090] Y 2DCNN =F 2DCNN (X2) (30)

[0091] Where: F 2DCNN Represents the frequency domain feature extraction module.

[0092] 3.2 Feature Fusion Layer

[0093] After flattening the feature vectors output by the time domain feature extraction module and the frequency feature extraction module, they are converted into two feature vectors of the same size through the fully connected layer. The two feature vectors are fused by splicing to obtain the fused feature Y cat , which is a local deep representation of the time-frequency characteristics of the harmonic source characteristics. The splicing operation can be expressed as follows:

[0094]

[0095] Where: Y cat It represents the fusion features after splicing, flatten is the flattening function, and Fc is the fully connected layer function.

[0096] 3.3 BiLSTM Network

[0097] The MCNN network can extract local characteristics of a short time period, but has certain deficiencies in extracting features of a long time period. In order to deeply explore the periodic change law of harmonic data on a long time scale, the present invention adopts a BiLSTM network to improve the timing performance of the model. The basic unit of the LSTM network is composed of a forget gate, an input gate, and an output gate. By selectively forgetting the information in the cell state and memorizing new information, it effectively transmits information that is useful for subsequent moment calculations, while discarding useless information. BiLSTM consists of two oppositely connected LSTMs, in which the forward LSTM is used to extract information before the sequence data, and the reverse LSTM is responsible for reverse processing of the data. The hidden layer fuses the forward and reverse information, and uses the past and future states to enhance the available information of the network, so as to accurately predict the extracted features. The input of the BiLSTM network is the fused feature Y obtained in Section 3.2 cat , the output of the model is the harmonic current X out , specifically:

[0098] X out =F BiLSTM (Y cat ) (32)

[0099] Where: F BiLSTM Represents a BiLSTM network.

[0100] 4 Harmonic source modeling method flow

[0101] (1) The voltage and current data of the harmonic source collected by the synchronous phasor measurement device are normalized and FFT analyzed. The content rate of each harmonic component is calculated, and the frequency of the main harmonic component of the harmonic source is determined by the threshold judgment rule specified in formula (4).

[0102] (2) Extract the harmonic voltage amplitude and phase angle corresponding to the harmonic main component frequency of the harmonic source to obtain the frequency domain feature set of the harmonic source. Reconstruct the frequency domain feature set through IDFT to form the frequency domain feature component matrix of the harmonic source, and splice it with the original voltage sequence collected by the synchronous phasor measurement device to obtain the time-frequency feature component matrix. Use the matrix normalization method to convert the component matrix into a grayscale image, and use the pseudo-color coding technology to convert it into a three-channel time-frequency feature component color image.

[0103] (3) Figure 1 As shown, the harmonic source voltage data collected by the synchronized phasor measurement device and the time-frequency characteristic component color map generated in the above steps are used as the input of the MCNN-BiLSTM model, and the corresponding harmonic source current data collected by the synchronized phasor measurement device are used as the output of the MCNN-BiLSTM model. The network model is continuously trained through a large amount of measured data to achieve accurate mapping of the relationship between the harmonic voltage and harmonic current of the harmonic source, and obtain the harmonic source data-driven model.

[0104] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0105] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0106] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0107] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0108] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0109] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0110] This patent is not limited to the above-mentioned optimal implementation mode. Anyone can derive other forms of a harmonic source modeling method based on time-frequency feature extraction under the inspiration of this patent. All equal changes and modifications made according to the scope of the patent application of the present invention should be covered by this patent.

Claims

1. A harmonic source modeling method based on time-frequency feature extraction, characterized in that: Firstly, the voltage and current data of the harmonic source collected by the synchronous phasor measurement device are normalized and analyzed by FFT; the content rate of each harmonic component is calculated, and the frequency of the main harmonic component of the harmonic source is determined by the set threshold judgment rule; Then, the harmonic voltage amplitude and phase angle corresponding to the harmonic main component frequency of the harmonic source are extracted to obtain the frequency domain feature set of the harmonic source; The frequency domain feature set is reconstructed by IDFT to form a harmonic source frequency domain feature component matrix, which is then spliced ​​with the original voltage sequence acquired by the synchronous phasor measurement device to obtain a time-frequency feature component matrix; the component matrix is ​​converted into a grayscale image using a matrix normalization method, and then converted into a three-channel time-frequency feature component color image using a pseudo-color coding technique; Finally, the harmonic source voltage data collected by the synchronized phasor measurement device and the time-frequency characteristic component color map are used as the input of the MCNN-BiLSTM model, and the corresponding harmonic source current data collected by the synchronized phasor measurement device are used as the output of the MCNN-BiLSTM model. The network model is continuously trained through measured data until the accurate mapping of the relationship between the harmonic voltage and harmonic current of the harmonic source is achieved, thereby obtaining a harmonic source data-driven model. The MCNN-BiLSTM model is a harmonic source voltage and current mapping model composed of a multiple convolutional neural network MCNN and a bidirectional long short-term memory network BiLSTM.

2. The harmonic source modeling method based on time-frequency feature extraction according to claim 1 is characterized in that: In the MCNN-BiLSTM model, the MCNN model extracts features jointly by a time domain feature extraction module and a frequency domain feature extraction module, and the two-way parallel structure realizes the simultaneous extraction of different feature information of multi-channel and multi-dimensional inputs; by serially connecting the MCNN and the BiLSTM, the MCNN model is extracted into the time-frequency domain fusion features and input into the BiLSTM network to obtain the feature information of the harmonic voltage in the entire time domain.

3. The harmonic source modeling method based on time-frequency feature extraction according to claim 1 is characterized in that: The specific steps of harmonic principal component analysis of harmonic sources are as follows: First, the raw data obtained by the synchronized phasor measurement device is normalized to unify the scale of the raw measurement data of the harmonic source: Where: is the normalized voltage and current sequence, U and I are the original voltage and current sequences measured by the synchronous phasor measurement device, and U max ,I max The extreme values ​​of the original voltage and current sequence measured by the synchronous phasor measurement device; Secondly, the normalized voltage and current sequence is subjected to fast Fourier transform to obtain the corresponding k-th order harmonic component: Where: U k , is the k-th order harmonic voltage amplitude and phase angle; I k , is the amplitude and phase angle of the k-th order harmonic current; Then, calculate the harmonic voltage content rate HRU k Harmonic current content HRI k : Where: U1 and I1 are the fundamental voltage and fundamental current amplitudes respectively; Finally, all HRUs are selected according to the following formula k and HRI k The harmonic frequencies that account for more than 1% are taken as the harmonic principal component frequencies, forming the harmonic principal component frequency set γ of the harmonic source; 4. The harmonic source modeling method based on time-frequency feature extraction according to claim 3 is characterized in that: According to the harmonic main component order set, the corresponding k-th harmonic voltage amplitude and phase angle are extracted to obtain the harmonic source frequency domain feature set. The steps to convert F into a time-frequency characteristic component color map are as follows: First, use the inverse Fourier transform to reconstruct the harmonic voltage components: Where: U k (t) represents the kth harmonic voltage characteristic component, N is the number of sampling points; Secondly, the harmonic voltage characteristic components {U1(t),U3(t),…,U k (t)} is concatenated with the original voltage sequence U(t) collected by the synchronous phasor measurement device to obtain the time-frequency characteristic component matrix {U1(t),U3(t),…,U k (t),U(t)}; Then, the time-frequency feature component matrix is ​​processed using the following formula, so that its data are all normalized to the range of (0, 255), and then it is matched with the grayscale value to obtain the grayscale image of the time-frequency feature; Where: G(m,n) represents the value of the gray matrix in the mth row and nth column; x i represents the value of the i-th data point; x min With x max Respectively represent the minimum and maximum values ​​in the time-frequency feature component matrix; Finally, the grayscale image is converted into a three-channel RGB color image using pseudo color coding to obtain a color map of the time-frequency feature components.

5. The harmonic source modeling method based on time-frequency feature extraction according to claim 2 is characterized in that: The time domain feature extraction module adopts a double-layer 1D convolutional network structure, each layer of the 1D convolutional network includes a 1D convolutional layer, a 1D regularization layer and a 1D pooling layer; the harmonic source voltage data X1 collected by the synchronized phasor measurement device is used as input, and passes through two layers of 1D convolutional networks in sequence, and finally obtains the feature vector Y 1DCNN : Y 1DCNN =F 1DCNN (X1) (7) Where: F 1DCNN Represents the time domain feature extraction module.

6. The harmonic source modeling method based on time-frequency feature extraction according to claim 2 is characterized in that: The frequency domain feature learning module adopts a two-layer 2D convolutional network, each layer of which includes a 2D convolutional layer, a 2D regularization layer, and a 2D pooling layer; the time-frequency feature component color map X2 is used as input, and passes through two layers of 2D convolutional networks in sequence, and finally obtains the feature vector Y 2DCNN : Y 2DCNN =F 2DCNN (X2) (8) Where: F 2DCNN Represents the frequency domain feature extraction module.

7. The harmonic source modeling method based on time-frequency feature extraction according to claim 2 is characterized in that: After the feature vectors output by the time domain feature extraction module and the frequency feature extraction module are flattened, they are converted into two feature vectors of the same size through the fully connected layer; the two feature vectors are fused by splicing to obtain the fused feature Y cat , as the local deep representation of the time-frequency features of the harmonic source characteristics, the splicing operation is expressed as follows: Where: Y cat It represents the fusion features after splicing, flatten is the flattening function, and Fc is the fully connected layer function.

8. The harmonic source modeling method based on time-frequency feature extraction according to claim 2 is characterized in that: The input of the BiLSTM network is the fusion feature Y cat , the output of the model is the harmonic current X out , specifically: X out =F BiLSTM (Y cat ) (11) Where: F BiLSTM Represents a BiLSTM network.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The computer program includes a harmonic source data-driven model obtained by a harmonic source modeling method based on time-frequency feature extraction as described in any one of claims 1-8.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: The computer program includes a harmonic source data-driven model obtained by a harmonic source modeling method based on time-frequency feature extraction as described in any one of claims 1-8.

Images