An attribute calibration method and system for alternating current-direct current power grid operation data

Abstract

This invention provides a method and system for attribute calibration of AC / DC power grid operation data. The calibration method includes: extracting instantaneous spectral information from AC / DC operation parameters using a smooth pseudo-affine Wigner-Wyler distribution and forming composite features; performing tree transformation on the composite features to generate nonlinear uncorrelated features; and learning the nonlinear uncorrelated features using a pre-constructed hierarchical Dirichlet process-remaining lifetime hidden semi-Markov model to achieve attribute calibration of the AC / DC power grid operation data corresponding to the operating state. The hierarchical Dirichlet process-remaining lifetime hidden semi-Markov model is based on adding the dwell time of abnormal states to the duration of the AC / DC power grid's operating state. This application, through the learning of the hierarchical Dirichlet process-remaining lifetime hidden semi-Markov model, can significantly detect abnormal states, solving a long-standing problem in AC / DC power grid safety assurance.

Description

Technical Field

[0001] This invention relates to the field of AC / DC power grid fault detection, and specifically to a method and system for attribute calibration of AC / DC power grid operation data. Background Technology

[0002] Learning about AC / DC grid faults is crucial for accurately assessing the state of AC / DC grids. This learning includes monitoring the performance of AC / DC grid operating parameters. Currently, the main learning modes for AC / DC grid operating parameters are: 1) learning the characteristics of changes in state variables such as AC-side unit power angle, node voltage, system frequency, and tie-line power under DC commutation failure and DC blocking conditions; 2) learning the characteristics of system power angle, voltage, and frequency under different instability modes after an AC / DC grid fault. Existing learning methods, such as traditional Fourier transform, cannot provide frequency domain transient information, and wavelet transform can only provide rich low-frequency information, making it difficult to provide three-dimensional synchronous display information of the signal's time, frequency, and amplitude. Hidden Markov Models (HMMs), as a classic model for hidden state learning, suffer from inconsistencies with practical applications due to the geometric distribution of state durations. Furthermore, HMMs require setting the number of hidden states; an inaccurate number will lead to deviations between the model and actual data. Bayesian non-parametric (BNP) methods are used to address this problem, inferring the correct number of states during the learning process. Applying BNP to Hidden Markov Models (HMMs) allows the number of hidden states to be inferred from the posterior distribution via a Hierarchical Dirichlet Process (HDP). Several extended models of HDP-HMM, such as adaptive HDP-HMM and online HDP-HMM, have been applied to handle relevant practical problems. Furthermore, traditional HDP-HMMs do not consider the duration of each state, leading to rapid state transitions. The problem of rapid state transitions is addressed by using Hidden Semi-Markov Models (HSMMs), which consider the dwell time of states. After entering a specific state, the Markov chain transitions to the next state when the dwell time ends. Unlike traditional HMMs, each state in an HSMM corresponds to a set of observations. However, in AC power grid fault state detection, the influence of fault signals on state parameters is very weak, making it difficult to detect the fault initiation point.

[0003] The search revealed the following: The literature "Research on the Impact of UHV AC / DC Access on Short-Circuit Current of Jiangxi Power Grid" only provides qualitative AC / DC analysis, offering measures and methods for addressing excessive short-circuit current, but without modeling or identifying abnormal AC / DC states; the literature "Fault Characteristic Analysis of Multi-Voltage Level AC / DC Hybrid Distribution Network Based on DAB DC Transformer" only conducts mechanism-based AC / DC power grid fault characteristic analysis, without transient analysis of AC / DC state parameters; and the literature "Empirical Analysis of the Impact of AC / DC Line Contact Faults on Transformer Differential Protection" simulates the AC / DC system, but the simulation results based on the system model are insufficient to verify real-world AC / DC fault data.

[0004] In summary, existing methods still fail to solve the problem of effectively calibrating the operational data attributes of AC and DC power grids during AC power grid fault learning. Summary of the Invention

[0005] To address the aforementioned shortcomings in the existing technology, this invention provides a method for attribute calibration of AC / DC power grid operation data, comprising:

[0006] Instantaneous spectral information of AC / DC operating parameters is extracted using a smooth pseudo-affine Wigner-Wyler distribution and combined into composite features.

[0007] The composite features are subjected to tree transformation to generate nonlinear, unrelated features;

[0008] The nonlinear uncorrelated features are learned by using a pre-constructed hierarchical Dirichlet process-remaining lifetime hidden semi-Markov model, thereby realizing the attribute labeling of AC / DC power grid operation data corresponding to the operating state.

[0009] The hierarchical Dirichlet process-remaining lifetime hidden semi-Markov model is modeled by adding the dwell time of abnormal states to the duration of AC / DC power grid operation.

[0010] Preferably, before extracting instantaneous spectral information from the AC / DC operating parameters using a smooth pseudo-affine Wigner-Wyler distribution and forming composite features, the method further includes:

[0011] The AC / DC operating parameters are denoised using wavelet packets.

[0012] The AC / DC operating parameters include: AC-side unit power angle, node voltage, and system frequency under DC commutation failure and DC blocking fault conditions.

[0013] Preferably, the extraction of instantaneous spectral information from AC / DC operating parameters using a smooth pseudo-affine Wigner-Wyler distribution includes:

[0014] The time-frequency distribution transformation of the AC / DC operating parameters is performed to obtain the Wigner distribution function;

[0015] By analyzing the signal in the Wigner distribution function, the Wigner-Wyler distribution function can be obtained;

[0016] Adding the Kaiser window function to the Wigner-Willer distribution function yields a smooth pseudo-Wigner-Willer distribution;

[0017] Instantaneous spectral information is extracted based on the time-frequency distribution of the smooth pseudo-Wigner-Wyler distribution.

[0018] Preferably, the AC / DC operating parameters are subjected to time-frequency distribution transformation according to the following formula:

[0019]

[0020] In the formula: s represents the AC / DC operating parameters; s * (u) is the conjugate complex number of s(u); It is the kernel function; u is the time variable; θ is... The function has fixed parameters; τ is the dwell time; ω is the frequency; and t is the time.

[0021] Preferably, the smooth pseudo-Wigner-Wyler distribution is as shown in the following formula:

[0022]

[0023] In the formula: W ′(l,m) For a smooth pseudo-Wigner-Wyler distribution; Δω is the frequency derivative; Δt is the time derivative, j is the time step; k is the frequency length; l is the time parameter of the frequency-time distribution; m is the angular frequency parameter of the frequency-time distribution; pl is the time t of the window function G; qm is the frequency ω of the window function G.

[0024] Preferably, the Kaiser window function is as shown in the following equation:

[0025]

[0026] In the formula: G(t,w) is the window function; τ is the dwell time of the Kaiser window; a is a non-negative real number that determines the shape of the window; I0 is the zero-order modified Bessel function of the first kind; t is time; ω is frequency.

[0027] Preferably, the step of performing a tree transformation on the composite features to generate nonlinear, uncorrelated features includes:

[0028] Calculate the covariance matrix and similarity measure for each AC / DC operating parameter feature in the composite features;

[0029] Find the covariance matrix with the highest similarity measure among all AC / DC operating parameter characteristics;

[0030] The found two-dimensional vector is subjected to principal component analysis transformation to obtain the Jacobian transformation matrix;

[0031] Based on the Jacobian transformation matrix, basis vectors with orthogonality are obtained for each layer, and nonlinear uncorrelated features are generated.

[0032] Preferably, the hidden semi-Markov model of the hierarchical Dirichlet process-remaining lifetime is as shown in the following equation:

[0033]

[0034] In the formula: f represents the AC / DC state characteristics, g represents the residence time distribution, GEM represents the folding rod construction process, γ represents the Gam function, d is the GAMMA distribution parameter, β is the rod construction process, α is the Dirichlet process parameter, iid is uniformly followed, and π j For the model scale, θ (i) For model parameters, Here are the residence time distribution parameters, H is the Dirichlet distribution, N is the normal distribution, and z s To enable the model to learn labels for AC and DC electrical characteristics, d s For the distribution of dwell time, y t To conform to the distribution The observation sequence, where G is a random measure. To concentrate on The probability measure, π k For random probability measures, z s For x s Indicator factor, c jk For the properties of the model, μ k Let K be the mean of the normal distribution, and K be the number of data components.

[0035] Based on the same inventive concept, the present invention also provides an AC / DC power grid operation data attribute calibration system, comprising:

[0036] The extraction module is used to extract instantaneous spectral information from AC / DC operating parameters using a smooth pseudo-affine Wigner-Wyler distribution and form composite features;

[0037] The transformation module is used to perform tree transformation on the composite features to generate nonlinear, unrelated features.

[0038] The learning module is used to learn the nonlinear uncorrelated features through a pre-constructed hierarchical Dirichlet process-remaining lifetime hidden semi-Markov model, so as to realize the attribute labeling of the AC / DC power grid operation data corresponding to the operation state.

[0039] The hierarchical Dirichlet process-remaining lifetime hidden semi-Markov model is based on modeling the duration of the AC / DC power grid's operating state.

[0040] Preferably, the transformation model includes:

[0041] A calculation unit is used to calculate the covariance matrix and similarity measure of each AC / DC operating parameter feature in the composite features;

[0042] The search unit is used to find the covariance matrix with the highest similarity measure among all AC and DC operating parameter characteristics;

[0043] The transformation unit is used to perform principal component analysis transformation on the found two-dimensional vector to obtain the Jacobian transformation matrix.

[0044] The generation unit is used to obtain the basis vectors with orthogonality in each layer based on the Jacobian transformation matrix, and to generate nonlinear uncorrelated features.

[0045] Preferably, the hidden semi-Markov model of the hierarchical Dirichlet process-remaining lifetime is as shown in the following equation:

[0046]

[0047] In the formula: f represents the AC / DC state characteristics, g represents the residence time distribution, GEM represents the folding rod construction process, γ represents the Gam function, c is the attribute label, d is the GAMMA distribution parameter, β is the rod construction process, α is the Dirichlet process parameter, iid represents uniformity, and π j For the model scale, θ (i) For model parameters, Here are the residence time distribution parameters, H is the Dirichlet distribution, N is the normal distribution, and z s To enable the model to learn labels for AC and DC electrical characteristics, d s For the distribution of dwell time, y t To conform to the distribution The observation sequence, where G is a random measure. To concentrate on The probability measure, π k For random probability measures, z s For x s Indicator factor, c jk For the properties of the model, μ k Let K be the mean of the normal distribution, and K be the number of data components.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] The technical solution provided by this invention extracts instantaneous spectral information from AC / DC operating parameters using a smooth pseudo-affine Wigner-Wyler distribution and forms composite features; performs tree transformation on the composite features to generate nonlinear uncorrelated features; learns the nonlinear uncorrelated features through a pre-constructed hierarchical Dirichlet process-remaining lifetime hidden semi-Markov model to achieve attribute labeling of AC / DC power grid operating data corresponding to operating states; the learning of the hierarchical Dirichlet process-remaining lifetime hidden semi-Markov model can significantly detect abnormal states, solving the long-standing problem of AC / DC power grid safety assurance.

[0050] This invention constructs a residual lifetime HSMM, which can characterize the degree of abnormality in the operating parameters of AC / DC power grids. It does not care about the starting point of the abnormality in AC / DC power grids, but rather the remaining time from the start of the abnormality to the extreme failure.

[0051] This invention employs a Hierarchical Dirichlet Process-Remaining Lifetime Hidden Semi-Markov Model (HDP-HSMM) to detect fault states in AC / DC power grids. It increases the dwell time of abnormal states, making faults easier to detect. Attached Figure Description

[0052] Figure 1 AC / DC power grid operation attribute calibration process

[0053] Figure 2 This refers to AC / DC power grid status information;

[0054] Figure 3 The characteristic energy of the time-frequency distribution spectrum obtained by SPAWVD;

[0055] Figure 4 The four nonlinear, uncorrelated feature distributions obtained by SPAWVD;

[0056] Figure 5 This is a schematic diagram of the HDP-HSMM structure. Detailed Implementation

[0057] To better understand this invention, the following description, in conjunction with the accompanying drawings and examples, will further illustrate the invention.

[0058] like Figure 1 As shown, the present invention provides a method for calibrating the operational data attributes of AC / DC power grids based on a hierarchical Dirichlet process-hidden semi-Markov model, comprising the following steps:

[0059] Step S1: Extract instantaneous spectral information from AC / DC operating parameters using a smooth pseudo-affine Wigner-Wyler distribution and construct composite features;

[0060] Step S2: Perform tree transformation on the composite features to generate nonlinear, unrelated features;

[0061] Step S3: The nonlinear uncorrelated features are learned by using a pre-constructed hierarchical Dirichlet process-remaining lifetime hidden semi-Markov model to realize the attribute labeling of the AC / DC power grid operation data corresponding to the operating state.

[0062] The hierarchical Dirichlet process-remaining lifetime hidden semi-Markov model is modeled by adding the dwell time of abnormal states to the duration of AC / DC power grid operation.

[0063] Furthermore, in step S1, since the AC / DC power grid state parameter signal is a mixed electrical signal, it is easily mixed with other noise during the data collection process. Therefore, the state parameter signal needs to be denoised first to reduce its impact on the subsequent feature extraction and recognition process. Figure 2 As shown, this application selects the db10 wavelet packet transform to decompose the typical state parameter signal, and then reconstructs the denoised typical state parameter signal from the wavelet coefficients.

[0064] To obtain the instantaneous spectral information of these AC / DC power grid operating parameter signals, an improved Wigner-Ville distribution (WVD) based on the Kaiser window is proposed. The Wigner distribution function (WDF) is an important method for analyzing non-stationary signals. Cross-term interference can cause negative values ​​in the spectral information. By incorporating a sliding exponential window to affine smooth the Wigner-Ville distribution, the interference of cross-terms can be effectively suppressed. This improved method is called Smoothed Pseudo-affine WVD (SPAWVD).

[0065] The time-frequency distribution of signal s(u) can be expressed as:

[0066]

[0067] Where s * (u) is the conjugate complex number of s(u). It is a kernel function.

[0068] Different distributions can be generated depending on the kernel function chosen. When we have this information, we can obtain the Wigner distribution:

[0069]

[0070] The energy spectral density function p(ω) can be expressed as:

[0071]

[0072] Where R t (τ) is the time-varying autocorrelation function, from which the time-varying energy spectral density function can be obtained.

[0073] For a continuous WDF:

[0074] w(t,ω)=∫R t (τ)e -jwτ dτ (4)

[0075] Where R t (τ) can be expressed in symmetric form as:

[0076]

[0077] Furthermore, the discrete-time WDF can be expressed as:

[0078]

[0079] When sampling a signal s(t), aliasing occurs in the Wigner-Willer distribution (WDF). An effective way to avoid aliasing is to analyze the signal before calculating the WDF. The WDF, also known as the Wigner-Willer distribution, can be represented as:

[0080]

[0081] Where H{s r The Hilbert transform (n) is generated by convolution of the impulse response h(t) with a 90-degree phase shift.

[0082] To avoid negative values ​​caused by interference terms, a Kaiser window function G(t,w) is added to WVD:

[0083]

[0084] Where τ is the dwell time of the Kaiser window, and a is a non-negative real number that determines the shape of the window. I0 is the zeroth-order modified Bessel function of the first kind.

[0085] Correspondingly, WVD can be represented as:

[0086]

[0087] By selecting appropriate w and t, the sampling Kaiser window function can be obtained:

[0088]

[0089] The values ​​of p and q vary between ±2j and ±2k, respectively. The smoothed pseudo-Wigner-Ville distribution (SPWVD) can be obtained by convolving the sampled WDF with the Kaiser window function.

[0090]

[0091] Then, the affine SPWVD can be represented as:

[0092]

[0093] Where Ψ(t,m) is a smooth function.

[0094] In this invention, the AC / DC power grid operating parameter s undergoes time-frequency distribution transformation according to equation (1), and then a three-dimensional time-frequency-amplitude spectrum of the power grid parameter s is obtained through equation (12). The spectrum density p is further obtained through SPAWVD transformation. s (ω). In this invention, the instantaneous spectrum of SPAWVD is used instead of the spectrum of FFT, because FFT can only provide average spectrum information within a given time period, while SPAWVD can provide the time-frequency amplitude information of AC / DC power grid operating status signals in real time.

[0095] To evaluate the advantages of SPAWVD time-frequency analysis, this invention provides four feature extraction methods: Wigner-Ville Distribution (WVD), Pseudo Wigner-Ville Distribution (PWVD), Smoothed Pseudo Wigner-Ville Distribution (SPWVD), and Smoothed Pseudo-affine Wigner-Ville Distribution (SPAWVD). These four extraction methods are used to process AC / DC power grid operating parameters, resulting in four different time-frequency distributions. For example... Figure 3 The time-frequency distribution spectrum features provided by SPAWVD show that PWVD has a stronger ability to extract local time-frequency features than WVD, and SPAWVD has a significantly stronger ability to represent local detail features than SPWVD. That is, both SPAWVD and SPWVD provide the instantaneous three-dimensional features (time, frequency domain, amplitude) of AC / DC power grid state parameter signals, but SPAWVD can extract richer instantaneous local spectrum information than SPWVD.

[0096] Compared with existing technologies, this invention proposes a smoothed pseudo-affine Wigner-Ville distribution (SPAWVD) to process the operating state variables of AC and DC power grids separately and obtain their instantaneous frequency, instantaneous amplitude, time-frequency distribution marginal spectrum and other characteristics.

[0097] Furthermore, in step 2, tree transforms (Treelets) are used to perform the following operations on the SPAWD-based time-spectrum features from step 1:

[0098] 1) Calculate the covariance matrix and similarity measure for each AC / DC parameter feature x;

[0099] 2) Find the covariance matrix with the highest similarity measure;

[0100] 3) Perform principal component analysis (PCA) on the found two-dimensional vector to find the Jacobian transformation matrix;

[0101] 4) After the Jacobian transformation, the basis vectors with orthogonality in each layer are obtained, the x data is reduced to k, and more than 90% of the original signal energy is retained.

[0102] from Figure 4 It can be seen that the time-spectral features learned by SPAWVD can be transformed by Treelet to obtain nonlinear and uncorrelated features, and these feature distributions have no overlap.

[0103] This invention proposes a Treelet similarity vector selection mechanism based on vector similarity metrics (maximum dissimilarity coefficient, relative error distance, generalized Dice coefficient, and generalized Jaccard coefficient), and establishes an optimal treelet transformation. The treelet transformation is used to perform a deassociation operation on the aforementioned features, generating linearly independent feature vectors. This allows for the merging of features with similar characteristics, thereby completing the preliminary feature extraction of AC / DC power grid operating status signals.

[0104] Furthermore, in step 3, the emission probability distribution function of the HDP-HSMM model is a Gaussian mixture distribution, the parameter sampling is a Dirichlet prior, and the state dwell probability follows a Possion distribution.

[0105] Build as follows: Figure 5 HDP-HSMM shown:

[0106] 1) To address the issue of having to set the number of hidden states, a Hierarchical Dirichlet Process (HDP) is added to HSMM as a prior for the infinite state space.

[0107] A Dirichlet process can be represented by a random probability density distribution G0 and a positive real number α0:

[0108] G~DP(α0,G0)

[0109] This invention uses a stick-breaking process to construct the Dirichlet process.

[0110]

[0111] Where G is a random measure, To concentrate on The probability measure, π k It is a measure of random probability.

[0112] The process of folding the stick can be recorded as follows:

[0113] π~GEM(α0)

[0114] 2) Similarly, the Dirichlet mixing process is constructed from the folding process:

[0115]

[0116] Where z s For x s Indicator factors.

[0117] Assume x ij Follows the distribution F(θ) ij ), θ ji Follows distribution G j And the value is The probability is π jk The hierarchical Dirichlet process can be represented as:

[0118]

[0119] 3) The generation process of HDP-HSMM is similar to that of HDP-HMM, and can be represented by the following formula:

[0120]

[0121] Where f represents the AC / DC state characteristics, g represents the residence time distribution, GEM represents the folding rod construction process, γ represents the Gam function, and z s To enable the model to learn labels for AC and DC electrical characteristics, d s For the distribution of dwell time, yt To conform to the distribution The observed sequence, c is the attribute label, d is the GAMMA distribution parameter, β is the stick construction process, α is the Dirichlet process parameter, iid is uniformity, π j For the model scale, θ (i) For model parameters, Here, H represents the residence time distribution parameter, N represents the Dirichlet distribution, and G represents the random measure. To concentrate on The probability measure, π k For random probability measures, z s For x s Indicator factor, c jk For the properties of the model, μ k Let K be the mean of the normal distribution, and K be the number of data components.

[0122] In this invention, the labels are in the form of 1, 2, and 3, which correspond to normal, slightly abnormal and fault states, respectively. The labels are a quantitative representation of attributes.

[0123] This invention constructs a Remaining Lifetime HSMM, which can characterize the degree of abnormal states in the operating parameters of AC / DC power grids. It does not concern itself with the starting point of the abnormal state, but rather with the remaining time (lifetime) from the onset of the abnormal state to a critical failure. Therefore, it can significantly detect abnormal states, which is precisely the problem that has been difficult to solve in ensuring the safety of AC / DC power grids, and is also the starting point for modeling.

[0124] Regarding the application of this model to detect AC / DC power grid fault states, this invention uses HDP-HSMM to detect AC / DC power grid fault states, which increases the dwell time of abnormal states, making faults easier to detect.

[0125] This invention proposes a multimodal feature extraction method for AC / DC power grid operating parameters based on advanced signal processing and a study on AC / DC power grid feature attribute identification based on Hierarchical Dirichlet Processes (HDP)-Remaining Useful Life (HSMM-RUL) Hidden semi-Markov Model with Remaining Useful Life (HSMM-RUL).

[0126] This invention provides a method for calibrating the operational data attributes of AC / DC power grids based on hierarchical Dirichlet processes and hidden semi-Markov models, solving the following technical problems:

[0127] (1) The typical state parameter signal is decomposed by wavelet packet transform of db10, and then the denoised typical state parameter signal is reconstructed by wavelet coefficients. The denoised state parameter signal is extracted by improved Wigner-Ville distribution (WVD) based on Kaiser window, which solves the problem that the feature data is too simple and noise interference is difficult to avoid.

[0128] (2) The optimal tree transformation is proposed to perform the correlation removal operation on the above features, generate linearly independent feature vectors, realize the merging of features with similar characteristics, thereby completing the preliminary feature extraction of AC / DC power grid operation status signals, and reducing the data dimension as much as possible while ensuring sufficient energy of the original signal.

[0129] (3) It provides a rigorous classification algorithm that is more adaptable to new data and is not entirely dependent on parameters. Furthermore, the classification algorithm should be able to handle non-linear data.

[0130] This embodiment analyzes three collected AC / DC operating parameters as examples. These parameters include several grid operating states: State 1 (normal), State 2 (minor anomaly state), and State 3 (fault state). For each operating parameter, SPAWVD time-frequency processing is performed to extract instantaneous spectral information, which is then combined into a composite feature. After Treelet transformation, four nonlinear uncorrelated features are formed. The latent variable z is then analyzed using the f method of HDP-HSMM. s Learn to calibrate the AC / DC operating status attributes.

[0131] Compared with other hidden state learning models such as Gaussian Mixture Models (GMM), Hidden Mixture Models (HMM), and HSMM, GMM is the weakest in processing AC / DC transient spectrum information. Clearly, Gaussian models are easily affected by data fluctuations. HMM, because it does not consider state dwell time, leads to rapid state transitions, resulting in lower recognition accuracy for anomalous states (micro-anomalies and fault states) compared to HSMM. The recognition accuracy of HSMM is influenced to some extent by the number of hidden states, and its accuracy in anomalous state recognition is significantly lower than HDP-HSMM. This is because HSMM uses the HDP method for parameter prior settings, thus HDP-HSMM is better able to accurately identify subtle features, which also verifies the advantage of HDP-HSMM in automatically inferring the number of hidden states. In real-world datasets, the SPAWVD-based feature extraction and HDP-HSMM state recognition methods proposed in this application both demonstrate high classification accuracy. Furthermore, Hidden Semi-Korkov Models (HSMs) are unsupervised learning models used for data labeling. Deep contraction models are a type of deep learning model that can perform both unsupervised feature learning and supervised recognition.

[0132] Based on the same inventive concept, the present invention also provides an AC / DC power grid operation data attribute calibration system, comprising:

[0133] The extraction module is used to extract instantaneous spectral information from AC / DC operating parameters using a smooth pseudo-affine Wigner-Wyler distribution and form composite features;

[0134] The transformation module is used to perform tree transformation on the composite features to generate nonlinear, unrelated features.

[0135] The learning module is used to learn the nonlinear uncorrelated features through a pre-constructed hierarchical Dirichlet process-remaining lifetime hidden semi-Markov model, so as to realize the attribute labeling of the AC / DC power grid operation data corresponding to the operation state.

[0136] The hierarchical Dirichlet process-remaining lifetime hidden semi-Markov model is based on modeling the duration of the AC / DC power grid's operating state.

[0137] In this embodiment, the transformation model includes:

[0138] A calculation unit is used to calculate the covariance matrix and similarity measure of each AC / DC operating parameter feature in the composite features;

[0139] The search unit is used to find the covariance matrix with the highest similarity measure among all AC and DC operating parameter characteristics;

[0140] The transformation unit is used to perform principal component analysis transformation on the found two-dimensional vector to obtain the Jacobian transformation matrix.

[0141] The generation unit is used to obtain the basis vectors with orthogonality in each layer based on the Jacobian transformation matrix, and to generate nonlinear uncorrelated features.

[0142] In this embodiment, the hidden semi-Markov model of the hierarchical Dirichlet process-remaining lifetime is shown in the following equation:

[0143]

[0144] In the formula: f represents the AC / DC state characteristics, g represents the residence time distribution, GEM represents the folding rod construction process, γ represents the Gam function, and z s To enable the model to learn labels for AC and DC electrical characteristics, d s For the distribution of dwell time, y t To conform to the distribution The observation sequence, where G is a random measure. To concentrate on The probability measure, π k For random probability measures, z s For xs Indicator factors.

[0145] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0146] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0147] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0148] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0149] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval.

Claims

1. An attribute calibration method for AC / DC grid operation data, characterized in that, include: Instantaneous spectral information of AC / DC operating parameters is extracted using a smooth pseudo-affine Wigner-Wyler distribution and combined into composite features. The composite features are subjected to tree transformation to generate nonlinear, unrelated features; The nonlinear uncorrelated features are learned by using a pre-constructed hierarchical Dirichlet process-remaining lifetime hidden semi-Markov model, thereby realizing the attribute labeling of AC / DC power grid operation data corresponding to the operating state. The hierarchical Dirichlet process-remaining lifetime hidden semi-Markov model is modeled by adding the dwell time of abnormal states to the duration of AC / DC grid operation; the AC / DC operating parameters include: AC side unit power angle, node voltage and system frequency under DC commutation failure and DC blocking fault.

2. The method of claim 1, wherein, Before extracting instantaneous spectral information from AC / DC operating parameters using a smooth pseudo-affine Wigner-Wyler distribution and forming composite features, the process also includes: The AC / DC operating parameters are denoised using wavelet packets.

3. The method of claim 1, wherein, The extraction of instantaneous spectral information from AC / DC operating parameters using a smooth pseudo-affine Wigner-Wyler distribution includes: The time-frequency distribution transformation of the AC / DC operating parameters is performed to obtain the Wigner distribution function; By analyzing the signal in the Wigner distribution function, the Wigner-Wyler distribution function can be obtained; Adding the Kaiser window function to the Wigner-Willer distribution function yields a smooth pseudo-Wigner-Willer distribution; Instantaneous spectral information is extracted based on the time-frequency distribution of the smooth pseudo-Wigner-Wyler distribution.

4. The method of claim 3, wherein, The AC / DC operating parameters are transformed in time-frequency distribution according to the following formula: In the formula: These are the AC / DC operating parameters; yes The conjugate of complex numbers; It is a kernel function; It is a time variable; for Fixed parameters of a function; For length of stay; For frequency; For time.

5. The method as described in claim 3, characterized in that, The smooth pseudo-Wigner-Wyler distribution is shown in the following equation: In the formula: It is a smooth pseudo-Wigner-Willer distribution; It is the frequency derivative; For time differentiation, For time step; Frequency; For time; For frequency; The time parameter of instantaneous frequency. The angular frequency parameter is the instantaneous frequency. For window functions time ; For window functions frequency .

6. The method as described in claim 3, characterized in that, The Kaiser window function is shown in the following equation: In the formula: For window functions; The dwell time at the Kaiser window; A nonnegative real number that determines the shape of the window; This is a zeroth-order modified Bessel function of the first kind; For time; For frequency.

7. The method as described in claim 1, characterized in that, The step of performing a tree transformation on the composite features to generate nonlinear, uncorrelated features includes: Calculate the covariance matrix and similarity measure for each AC / DC operating parameter feature in the composite features; Find the covariance matrix with the highest similarity measure among all AC / DC operating parameter characteristics; The found two-dimensional vector is subjected to principal component analysis transformation to obtain the Jacobian transformation matrix; Based on the Jacobian transformation matrix, basis vectors with orthogonality are obtained for each layer, and nonlinear uncorrelated features are generated.

8. The method as described in claim 1, characterized in that, The hidden semi-Markov model of the hierarchical Dirichlet process-remaining lifetime is shown in the following equation: In the formula: It is the distribution function of observed values, characterizing the AC / DC state features. The GEM represents the residence time distribution, and the GEM represents the folding structure process. Represents the Gam function. The shape parameter of the GAMMA distribution. For GAMMA distribution parameters, This describes the process of constructing a folding rod. For Dirichlet process parameters, Hidden state j The transition probability distribution, For hidden state sequence The transition probability distribution, Hidden state i Model parameters, Hidden state i The residence time distribution parameters, The index number for the hidden state. For a normal distribution, use Indicates the distribution of stay time , To conform to the distribution The observation sequence, This indicates the identifier of the hidden state at observation time t. This represents the set of hidden state labels corresponding to all observation times within the time interval [t(s), t(s+1)-1]. For random probability measure, The mean of a normal distribution is . This represents the number of data components.

9. A system for calibrating the operational data attributes of an AC / DC power grid, characterized in that, include: The extraction module is used to extract instantaneous spectral information from AC / DC operating parameters using a smooth pseudo-affine Wigner-Wyler distribution and form composite features; The transformation module is used to perform tree transformation on the composite features to generate nonlinear, unrelated features. The learning module is used to learn the nonlinear uncorrelated features through a pre-constructed hierarchical Dirichlet process-remaining lifetime hidden semi-Markov model, so as to realize the attribute labeling of the AC / DC power grid operation data corresponding to the operation state. The hierarchical Dirichlet process-remaining lifetime hidden semi-Markov model is based on modeling the duration of the AC / DC power grid operating state; the AC / DC operating parameters include: AC side unit power angle, node voltage and system frequency under DC commutation failure and DC blocking fault.

10. The system as described in claim 9, characterized in that, The transformation module includes: A calculation unit is used to calculate the covariance matrix and similarity measure of each AC / DC operating parameter feature in the composite features; The search unit is used to find the covariance matrix with the highest similarity measure among all AC and DC operating parameter characteristics; The transformation unit is used to perform principal component analysis transformation on the found two-dimensional vector to obtain the Jacobian transformation matrix. The generation unit is used to obtain the basis vectors with orthogonality in each layer based on the Jacobian transformation matrix, and to generate nonlinear uncorrelated features.

11. The system as described in claim 9, characterized in that, The hidden semi-Markov model of the hierarchical Dirichlet process-remaining lifetime is shown in the following equation: In the formula: It is the distribution function of observed values, characterizing the AC / DC state features. The GEM represents the residence time distribution, and the GEM represents the folding structure process. Represents the Gam function. The shape parameter of the GAMMA distribution. For GAMMA distribution parameters, This describes the process of constructing a folding rod. For Dirichlet process parameters, Hidden state j The transition probability distribution, For hidden state sequence The transition probability distribution, Hidden state i Model parameters, Hidden state i The residence time distribution parameters, The index number for the hidden state. For a normal distribution, use Indicates the distribution of stay time , To conform to the distribution The observation sequence, This indicates the identifier of the hidden state at observation time t. This represents the set of hidden state labels corresponding to all observation times within the time interval [t(s), t(s+1)-1]. For random probability measure, The mean of a normal distribution is . This represents the number of data components.

Images