A power quality composite disturbance co-modeling method based on multi-source data

Abstract

The present application relates to the technical field of electric digital data processing, in particular to a power quality composite disturbance collaborative modeling method based on multi-source data, comprising: constructing an over-complete atom library based on distributed node collection sequence characteristics; projecting multi-source power grid sampling sequences to the atom library, identifying and stripping the dominant electric energy component with the highest energy contribution degree; stripping the output residual sequence through recursive subtraction until the convergence criterion of normalized energy proportion is met; detecting residual pulse feature vectors from the residual sequence; performing spatial coherence calculation on the residual pulse feature vectors using a preset propagation time delay compensation window, checking spatial dimension redundancy through sliding window search to eliminate false features; and performing parameter alignment and nonlinear reorganization on the effective disturbance component and the background feature set to generate a composite disturbance feature matrix. The present application solves the problem that weak disturbance components are difficult to identify under strong background energy shielding, and realizes high-fidelity collaborative modeling of composite disturbance patterns in a heterogeneous data environment.

Description

Technical Field

[0001] This invention relates to the field of electrical digital data processing technology, specifically to a collaborative modeling method for power quality composite disturbances based on multi-source data. Background Technology

[0002] In the context of real-time processing of large-scale heterogeneous data, the non-stationary sequences generated in the power system monitoring field, especially in modern distribution networks with a high proportion of fluctuating renewable energy grid connection and large-scale access of power electronic converters, exhibit extremely high computational complexity. For the automated identification of power quality signals, traditional computational frameworks are typically deployed within digital signal analysis or computer vision architectures, utilizing discrete Fourier transform, multi-scale wavelet transform, or discrete transform operators to extract and map feature vectors. The core logic of this modeling method lies in using a pre-defined orthogonal basis function space to perform projection operations on the original discrete sampling sequence. By calculating the amplitude, phase, or time-frequency distribution of feature operators, mathematical vectors reflecting the morphological characteristics of the power frequency fundamental wave, harmonics, or transient disturbances are extracted, providing deterministic algorithmic input for subsequent power big data analysis.

[0003] However, existing computational models suffer from significant shortcomings in feature space decoupling and numerical fidelity when processing complex electrical energy perturbation sequences composed of multiple nonlinear functions. Specifically, limited by the computational constraints of fixed orthogonal basis, severe feature energy leakage occurs in the transform domain when high-energy dominant components and weak pulse feature components in the signal undergo nonlinear superposition. This causes the feature operators of weak components to be completely masked by the computational sidelobe noise of the high-energy dominant components during quantization, resulting in information topology distortion and mode mixing in the feature space. Furthermore, existing algorithms lack adaptive iterative stripping logic for incoherent heterogeneous components, making it impossible to achieve orthogonal decomposition of each physical component under conditions of extreme energy imbalance, directly limiting the analytical accuracy of recognition models for complex composite patterns. Therefore, designing a data processing architecture with adaptive feature stripping capabilities to achieve sparse decomposition and accurate reconstruction of weak useful components under strong background energy masking is a key technical problem that urgently needs to be solved in the fields of computer data processing and signal processing.

[0004] To address this, a collaborative modeling method for power quality composite disturbances based on multi-source data is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a collaborative modeling method for power quality composite disturbances based on multi-source data. By deeply coordinating time-frequency atomic deconstruction and distributed spatial coherence verification, it solves the problem of difficulty in identifying weak composite disturbance components under strong background energy masking, and realizes high-fidelity modeling of composite disturbance modes in heterogeneous data environments.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A collaborative modeling method for power quality composite disturbances based on multi-source data includes:

[0008] Based on the attenuation and oscillation time-frequency characteristics of multi-source power grid sampling sequences collected by distributed nodes, an overcomplete atomic library containing a set of primitive functions is constructed.

[0009] The multi-source power grid sampling sequence is projected onto the overcomplete atom library to perform projection operator operations. By calculating the inner product of each atom in the overcomplete atom library with the multi-source power grid sampling sequence, the main conductive energy component with the highest energy contribution is identified and output.

[0010] The main conductive power component is recursively subtracted from the multi-source power grid sampling sequence to output a residual sequence;

[0011] The residual sequence is used as the input for the next iteration. The projection operator is returned to perform cyclic stripping until the energy magnitude of the residual sequence meets the preset convergence criterion. The residual pulse feature vector is then detected and output from the residual sequence.

[0012] The residual pulse feature vector is combined with the time-frequency atomic distribution of adjacent nodes in the distributed nodes, and coherence calculation is performed. By verifying the redundancy of the residual pulse feature vector in the spatial dimension, false features generated by random noise are eliminated, and a composite perturbation feature matrix is ​​reconstructed.

[0013] Preferably, the process of constructing an overcomplete atomic library based on the distributed nodes includes: performing analog-to-digital conversion on the original power signal using the distributed nodes deployed at the distribution feeder end, and outputting the multi-source power grid sampling sequence; determining a disturbance physical parameter space including the envelope attenuation coefficient range and the oscillation center frequency range based on the statistical characteristics of the multi-source power grid sampling sequence; injecting the attenuation factor distribution, frequency operator distribution, and displacement factor distribution generated by mapping the disturbance physical parameter space as parameters to be optimized into a preset primitive function set; performing multi-scale discretization iterative sampling on the primitive function set to generate an atom set covering different time-frequency granularities, and constructing the overcomplete atomic library.

[0014] Preferably, projecting the multi-source power grid sampling sequence onto the overcomplete atom library to perform projection operator operations includes: performing conjugate transpose multiplication on the multi-source power grid sampling sequence and each atom in the atom set of the overcomplete atom library to output a sequence of inner product real numbers reflecting the degree of waveform matching; performing absolute value processing on the inner product real number sequence to generate an inner product modulus distribution vector; searching for the global maximum value in the inner product modulus distribution vector, and labeling the index value corresponding to the global maximum value as the optimal atom index of the current iteration.

[0015] Preferably, identifying and outputting the main conductive energy component with the highest energy contribution includes: retrieving the corresponding target primitive waveform function from the overcomplete atom library according to the optimal atom index; performing a scalar multiplication operation between the target primitive waveform function and the weight coefficient at the corresponding index position in the inner product real number sequence to generate the main conductive energy component.

[0016] Preferably, the process of performing the recursive subtraction stripping and outputting the residual sequence includes: taking the multi-source power grid sampling sequence as the signal sequence to be processed in the current round, and taking the main conductive energy component as the deduction item in the current round; performing point-by-point numerical subtraction operation on the signal sequence to be processed and the deduction item using the subtraction operator, and outputting the temporary residual vector of the current round; and labeling the temporary residual vector as the residual sequence.

[0017] Preferably, the residual sequence is used as the input for the next iteration to perform loop stripping, including: determining whether the energy magnitude of the residual sequence is higher than the preset convergence criterion; if the determination result is higher than the preset convergence criterion, then a recursive call instruction is initiated to map the residual sequence output in the current round to the signal sequence to be processed in the next iteration loop, and the projection operator operation is returned to be executed; if the determination result is not higher than the preset convergence criterion, then the iteration loop is terminated and the final residual sequence is locked, and the final residual sequence is input to the feature detection operator to generate the residual pulse feature vector.

[0018] Preferably, the convergence criterion is met until the energy modulus of the residual sequence satisfies the preset convergence criterion, including: calculating the square of the L2 norm of the residual sequence output in the current iteration to obtain the current energy modulus; calculating the ratio of the current energy modulus to the total energy modulus of the multi-source power grid sampling sequence to output the normalized energy ratio; performing a comparison operation between the normalized energy ratio and a preset calculation accuracy limit; if the normalized energy ratio is lower than the calculation accuracy limit, it is determined that the convergence criterion is met.

[0019] Preferably, detecting and outputting the residual pulse feature vector from the residual sequence includes: using a high-pass filter operator to filter out low-frequency residual components in the residual sequence and outputting a high-frequency component sequence; performing time-domain segmentation mapping on the high-frequency component sequence to extract a set of sampling points whose amplitude jump rate exceeds a preset rate of change benchmark; using a pulse positioning operator to determine the distribution range of the sampling point set on the time axis and outputting a pulse position index; and based on the pulse position index, extracting the corresponding numerical components from the residual sequence to construct and generate the residual pulse feature vector.

[0020] Preferably, the process of performing the coherence calculation and removing pseudo-features generated by random noise includes: obtaining the time-frequency atom distributions of each distributed node physically adjacent to the current node within a preset propagation delay compensation window; performing a sliding window search on the residual pulse feature vector using a cross-correlation operator, and performing a dot product operation with each of the time-frequency atom distributions respectively, outputting a spatial coherence score sequence; labeling the values ​​in the spatial coherence score sequence below a preset consistency threshold as pseudo-feature components, and labeling the remaining values ​​that meet the preset consistency threshold as valid perturbation components.

[0021] Preferably, the process of reconstructing and generating the composite perturbation feature matrix includes: extracting the amplitude envelope parameters and center frequency parameters of the main conductive energy component output in each iteration to construct a background feature set; mapping the effective perturbation component to the coordinate space where the background feature set is located according to the sampling timestamp index; calculating the coupling weight between the effective perturbation component and the background feature set; performing parameter alignment and nonlinear recombination processing to generate a cooperative feature vector; and mapping the cooperative feature vector to a matrix operator template according to a preset topological dimension, wherein, based on the difference between the actual number of adjacent nodes and the preset column dimension, adaptive filling or filtering operations are performed on the cooperative feature columns in the matrix operator template to output the composite perturbation feature matrix.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] 1. By constructing an overcomplete atomic library containing multi-scale decay and oscillatory primitives, this invention utilizes projection operator operations and a recursive subtraction stripping mechanism to eliminate the dominant energy components layer by layer, suppressing the energy leakage and sidelobe masking effects generated by traditional fixed orthogonal transformations. This allows the low-energy pulse characteristics remaining in the residual sequence to be revealed, enhancing the ability to capture weak useful components in composite perturbations.

[0024] 2. By introducing the normalized energy ratio of the square of the L2 norm as a convergence criterion, this invention reduces the dimensional impact of different voltage levels on the calculation accuracy, ensuring the convergence of the algorithm in a heterogeneous node environment; combined with high-pass filtering and pulse positioning operators, it effectively eliminates the interference of harmonic residues on pulse components, ensuring the accuracy of residual pulse feature vector extraction.

[0025] 3. By performing cross-node cross-correlation sliding window search within a preset propagation delay compensation window, this invention can effectively identify and eliminate isolated spurious features caused by sensor random noise, retaining only effective perturbation components with physical coherence. Combined with the composite perturbation feature matrix generated by nonlinear recombination logic, it can more realistically characterize the physical coupling relationship between the dominant background and transient pulses, providing a high-fidelity data model for the refined management of power big data. Attached Figure Description

[0026] Figure 1 This is a flowchart of a collaborative modeling method for power quality composite disturbances based on multi-source data according to the present invention.

[0027] Figure 2 This is a flowchart of the iterative stripping process according to an embodiment of the present invention;

[0028] Figure 3 This is a flowchart of spatial coherence verification and pseudo-feature removal in an embodiment of the present invention. Detailed Implementation

[0029] The specific embodiments disclosed in this specification are intended to exemplify the technical logic of the present invention, and not to exhaustively limit the scope of protection. For those skilled in the art, all equivalent technical means or logical extensions derived from this disclosure without additional inventive effort should be considered to be covered within the scope of the claims of this invention.

[0030] Please see Figures 1 to 3 This invention provides a collaborative modeling method for power quality composite disturbances based on multi-source data, the technical solution of which is as follows:

[0031] A collaborative modeling method for power quality composite disturbances based on multi-source data includes:

[0032] Based on the attenuation and oscillation time-frequency characteristics of multi-source power grid sampling sequences collected by distributed nodes, an overcomplete atomic library containing a set of primitive functions is constructed.

[0033] The multi-source power grid sampling sequence is projected onto the overcomplete atom library to perform projection operator operations. By calculating the inner product of each atom in the overcomplete atom library with the multi-source power grid sampling sequence, the main conductive energy component with the highest energy contribution is identified and output.

[0034] The main conductive power component is recursively subtracted from the multi-source power grid sampling sequence to output a residual sequence;

[0035] The residual sequence is used as the input for the next iteration. The projection operator is returned to perform cyclic stripping until the energy magnitude of the residual sequence meets the preset convergence criterion. The residual pulse feature vector is then detected and output from the residual sequence.

[0036] The residual pulse feature vector is combined with the time-frequency atomic distribution of adjacent nodes in the distributed nodes, and coherence calculation is performed. By verifying the redundancy of the residual pulse feature vector in the spatial dimension, false features generated by random noise are eliminated, and a composite perturbation feature matrix is ​​reconstructed.

[0037] Example 1:

[0038] This embodiment takes the processing of transient harmonic composite disturbance data collected by the distributed power quality monitoring terminal (FTU / DTU) of the distribution network as the background.

[0039] First, the process of constructing an overcomplete atomic library based on the distributed nodes includes: performing analog-to-digital conversion on the original power signal using the distributed nodes deployed at the power distribution feeder end, and outputting the multi-source power grid sampling sequence; determining a disturbance physical parameter space containing the envelope attenuation coefficient range and the oscillation center frequency range based on the statistical characteristics of the multi-source power grid sampling sequence; injecting the attenuation factor distribution, frequency operator distribution, and displacement factor distribution generated by mapping the disturbance physical parameter space as parameters to be optimized into a preset primitive function set; performing multi-scale discretization iterative sampling on the primitive function set to generate an atomic set covering different time-frequency granularities, and constructing the overcomplete atomic library.

[0040] Specifically, in this embodiment, the distributed nodes adopt power quality monitoring instruments with synchronous sampling function, and their core processors adopt embedded digital signal microprocessors or field programmable gate array chips.

[0041] The distributed nodes perform analog-to-digital conversion on the original power signal. Specifically, the sampling frequency is set to 12.8 kHz, and the voltage signal of the 10 kV distribution network line is discretized and sampled. The discrete multi-source power grid sampling sequence is obtained through a 16-bit analog-to-digital converter.

[0042] Based on the statistical characteristics of the multi-source power grid sampling sequence, a disturbance physical parameter space including the range of envelope attenuation coefficient and the range of oscillation center frequency is determined. Transient fluctuation intervals of the signal are identified by performing energy centroid detection on the sampling sequence. Based on the duration of the fluctuation interval and the physical criterion that the envelope attenuates to 5% of its initial amplitude at the end of the time window, the search range for the attenuation factor is calibrated to be 0.01 to 100. The initial frequency is estimated by calculating the median time interval between adjacent zero-crossing points, and the search range for the center frequency is set to be 50 Hz to 3000 Hz.

[0043] The attenuation factor distribution, frequency operator distribution, and displacement factor distribution generated by the spatial mapping of the disturbance physical parameters are injected into a preset set of basic functions as parameters to be optimized. By taking the logarithm base 2 of the search interval 0.01 to 100, an exponential sequence of integers between 0 and 14 is obtained. This exponential sequence is used as the discrete order to obtain a set of attenuation factors composed of the products of 0.01 and 2 raised to their respective integer powers. The set of basic functions consists of the product of an attenuation envelope term and a periodic oscillation term. The attenuation envelope term uses an exponential attenuation function with the natural constant as the base and the product of the negative attenuation factor and the time deviation as the exponent. The periodic oscillation term uses a sine or cosine trigonometric function with the oscillation frequency, time deviation, and a preset initial phase as independent variables. In this embodiment, the initial phase is a fixed value of zero or half of the initial phase, and the time deviation is determined by the displacement factor.

[0044] Preferably, in order to enable those skilled in the art to reproduce overcomplete atomic library lookup tables with dimensions between 1024 and 4096, this embodiment further quantifies the discretization sampling distribution criteria of each operator as follows: in the frequency band of 50 Hz to 500 Hz, a fixed sampling step size of 2 Hz is set; in the high frequency band of 500 Hz to 3000 Hz, a percentage step size with a scaling factor of 0.05 is adopted, that is, the step size increases by 5% of the current center frequency, for example, the step size at 1000 Hz is 50 Hz; in the range of 0.01 to 100, iterative sampling is performed according to the product of the initial value 0.01 and an integer power of 2 until the upper limit of the coverage interval is reached; based on the perturbation trigger time, discrete displacement is performed in the feature observation window before and after the perturbation trigger time, with the sampling period as the step size.

[0045] Multi-scale discretization iterative sampling is performed on the primitive function set to generate an atom set covering different time-frequency granularities. This embodiment pre-prunes the atom library by introducing typical perturbation templates defined based on internationally accepted standards: the typical perturbation templates include a voltage dip template and a transient oscillation template; the voltage dip template uses the product of a square wave envelope with an amplitude reduced to 0.1 to 0.9 times the nominal value and a power frequency sine wave; the transient oscillation template uses the product of a sine wave with a fixed frequency of 500 Hz to 1500 Hz and a rapidly decaying exponential envelope. The correlation between candidate atoms generated by each parameter combination and the standard template is calculated. Only atoms with a correlation coefficient higher than 0.15 are retained. Through this screening mechanism, physically rare parameter combinations are eliminated, reducing the final atom library size stored in the monitoring terminal's random access memory to approximately 2048 atoms, resulting in the overcomplete atom library.

[0046] This invention constructs a highly matched, overcomplete atomic library through multi-scale discretization sampling and a typical perturbation template pre-pruning mechanism. Compared to blind global search, this method utilizes signal statistical characteristics to limit the physical parameter space, reducing the redundancy of the atomic library. The simplified library size effectively reduces the inner product computation load of the microprocessor, improving the real-time performance and engineering reproducibility of the distributed terminal execution algorithm while ensuring the accuracy of capturing transient harmonic composite perturbations.

[0047] Further, projecting the multi-source power grid sampling sequence onto the overcomplete atom library to perform projection operator operations includes: performing conjugate transpose multiplication on the multi-source power grid sampling sequence and each atom in the atom set of the overcomplete atom library to output an inner product real number sequence reflecting the degree of waveform matching; performing absolute value processing on the inner product real number sequence to generate an inner product modulus distribution vector; searching for the global maximum value in the inner product modulus distribution vector, and labeling the index value corresponding to the global maximum value as the optimal atom index of the current iteration.

[0048] Specifically, in this embodiment, the multi-source power grid sampling sequence to be processed is a discrete voltage vector containing 512 sampling points. First, the multi-source power grid sampling sequence is input to the multiply-accumulate unit of the digital signal microprocessor. In this embodiment, since both the sampling sequence and the atom set in the overcomplete atom library are in real number format, the "conjugate transpose multiplication" is simplified to a dot product operation of real number vectors. The 2048 atom vectors are arranged with consecutive addresses in random access memory, and the multiply-accumulate operation is performed simultaneously on four or eight sampling points using the microprocessor's single instruction multiple data stream instruction set.

[0049] Since each atom in the overcomplete atom library has been pre-normalized (i.e., the square of the L2 norm of each atom is 1), the dot product of the signal and the atom is calculated to directly output a sequence of real inner products reflecting the degree of waveform matching. If the size of the overcomplete atom library is 2048, a vector containing 2048 inner product values ​​is generated, where the magnitude of each value represents the similarity between the corresponding atom and the current perturbation signal.

[0050] Subsequently, the sequence of real inner products is subjected to absolute value conversion to generate an inner product modulus distribution vector. This process aims to reduce the polarity effect caused by the initial phase shift, focusing only on the overlap of perturbation energies. For example, if the inner product value of a certain atom is -0.85, it becomes 0.85 after absolute value conversion, ensuring that strong perturbation components with opposite phases but consistent morphology can be captured in subsequent optimization processes.

[0051] Next, the global maximum value in the inner product modulus distribution vector is searched. To address potential conflicts between candidate atoms with very similar values ​​or multiple peaks that may occur under complex perturbations, this embodiment adopts a uniqueness criterion of "value priority, minimum index": when traversing 2048 inner product modulus values, if the difference between the modulus values ​​at two different index positions is found to be less than a preset discrimination threshold (e.g., 0.001), the atom with the smaller index value (i.e., ranked higher in the overcomplete atom library) is automatically locked as the optimal atom index for the current iteration, thus avoiding oscillations between local extrema in the algorithm.

[0052] This invention utilizes conjugate transpose multiplication and absolute value conversion to achieve rapid similarity mapping between signals and the atom library through dot product operations, effectively reducing the impact of phase offset on matching results. By introducing a uniqueness criterion of "numerical priority, minimum index," combined with convergence criteria to perform recursive stripping, it can systematically extract the main conductive energy components and reduce residual energy. This method improves the accuracy of optimal atom index location while reducing local extremum oscillations, providing a stable algorithmic guarantee for the layer-by-layer deconstruction of complex perturbation features.

[0053] Furthermore, identifying and outputting the main conductive energy component with the highest energy contribution includes: retrieving the corresponding target primitive waveform function from the overcomplete atom library according to the optimal atom index; performing a scalar multiplication operation between the target primitive waveform function and the weight coefficient at the corresponding index position in the inner product real number sequence to generate the main conductive energy component.

[0054] Specifically, after determining the optimal atom index for the current iteration (e.g., index number 156), the digital signal microprocessor retrieves the corresponding target primitive waveform function from a pre-stored overcomplete atom library lookup table based on this index address. In this embodiment, since the atoms in the overcomplete atom library are discrete waveforms that have undergone pre-normalization, their amplitudes do not represent the actual voltage intensity.

[0055] Therefore, a scalar multiplication operation is performed between the retrieved target primitive waveform function and the weighting coefficients at the corresponding index positions in the aforementioned inner product real number sequence. Specifically, the weighting coefficients are the signed original values ​​of the inner product output in the aforementioned projection operation (in this embodiment, their physical meaning represents the projection length of the signal in the direction of that atom, i.e., the peak amplitude of the perturbation component). For example, if the real value of the inner product at the current index position is 220.5, then the scalar multiplication operator uses this value as a gain coefficient to proportionally scale the coordinates of each sampling point in the target primitive waveform function.

[0056] The above calculations generate the main conductive energy component. This component is a discrete numerical sequence with the same dimension as the original sampling sequence (i.e., 512 sampling points), and physically reconstructs the power quality disturbance component with the largest energy contribution at that moment. It is worth noting that, due to the use of scalar multiplication, the reconstructed main conductive energy component fully preserves the time-frequency characteristics of the target primitive, while also endowing it with the amplitude attribute that truly reflects the grid voltage fluctuations.

[0057] This invention achieves accurate reconstruction of the main conduction energy component through scalar multiplication of optimal atom indices and weighting coefficients. By using the original value of the inner product as a gain coefficient to proportionally scale the normalized primitive waveform, the true amplitude properties of the perturbation component can be effectively restored while fully preserving the time-frequency characteristics of the target primitive. This method ensures the consistency of the reconstructed sequence with the original sampling sequence in terms of physical dimensions, providing a reliable data foundation for subsequent precise recursive stripping and deep deconstruction of complex perturbation components.

[0058] Further, the process of performing the recursive subtraction stripping and outputting the residual sequence includes: taking the multi-source power grid sampling sequence as the signal sequence to be processed in the current round, and taking the main conductive energy component as the deduction item in the current round; using the subtraction operator to perform point-by-point numerical subtraction operation on the signal sequence to be processed and the deduction item, and outputting the temporary residual vector of the current round; and labeling the temporary residual vector as the residual sequence.

[0059] Specifically, at the start of the first iteration, the digital signal microprocessor labels the originally acquired multi-source power grid sampling sequence (such as a voltage amplitude vector containing 512 sampling points) as the signal sequence to be processed in the current iteration. Simultaneously, the main conductive energy component reconstructed in the aforementioned steps is labeled as the deduction term for the current iteration.

[0060] During the stripping operation, the subtraction operator performs a point-by-point numerical subtraction operation between the signal sequence to be processed and the deduction item. Specifically, it iterates through each discrete point index (from 1 to 512) on the sampling time axis, reads the voltage value of the signal sequence to be processed at that point, and reads the reconstructed value of the main conductive energy component at the same position. The difference between the two (voltage value minus reconstructed value) is calculated, and the energy component in the multi-source power grid sampling sequence that matches the current optimal atomic characteristics is removed.

[0061] By performing a subtraction traversal on all 512 sampling points, a temporary residual vector for the current round is output. Subsequently, this temporary residual vector is labeled as the residual sequence. In this embodiment, since the main conductive energy component is reconstructed based on projection weights, the energy of the stripped residual sequence in the corresponding perturbation frequency band will significantly decrease, thereby allowing weak pulse components or minor perturbation features masked by the high-energy background in the original signal to stand out in the residual space.

[0062] After completing the point-by-point subtraction operation for the aforementioned 512 sampling points, the digital signal microprocessor directly writes the resulting temporary residual vector back to the initial buffer storing the multi-source power grid sampling sequence to update the signal sequence to be processed. For detailed logic regarding how the updated signal triggers the next iteration and determines termination, please refer to the following description.

[0063] This invention achieves precise decoupling of signal components through point-by-point numerical subtraction. By using the subtraction operator to remove the dominant component of the current round, weak pulses or minor perturbation features that were originally masked by the high-energy background can be highlighted in the residual sequence. This method optimizes the computation process through a direct write-back buffer strategy, which not only reduces the residual energy but also provides a more targeted sequence to be processed for subsequent iterations and accurate detection of weak features, thus helping to improve the sensitivity of complex perturbation analysis.

[0064] Furthermore, referring to Figure 2 The residual sequence is used as the input for the next iteration to perform loop stripping, including: determining whether the energy magnitude of the residual sequence is higher than the preset convergence criterion; if the determination result is higher than the preset convergence criterion, a recursive call instruction is initiated to map the residual sequence output in the current round to the signal sequence to be processed in the next iteration loop, and the projection operator operation is returned to be executed; if the determination result is not higher than the preset convergence criterion, the iteration loop is terminated and the final residual sequence is locked, and the final residual sequence is input to the feature detection operator to generate the residual pulse feature vector.

[0065] Specifically, in this embodiment, a logical judgment mechanism is used to achieve layered stripping of complex perturbation components: after each round of subtraction stripping operation, the digital signal microprocessor reads the residual sequence in the current iteration buffer. A termination judgment with dual safeguards is then executed, simultaneously monitoring the energy percentage and the iteration counter.

[0066] Under typical operating conditions in power distribution networks, the dominant components of complex disturbances (such as voltage dips, low-order harmonics, and transient oscillations) usually exhibit significant energy sparsity, with the number of their independent components rarely exceeding 10 in physical and logical terms. Therefore, setting 10 as the upper limit can cover most extreme complex operating conditions while effectively filtering out high-order nonlinear noise not covered by the atomic library, preventing excessive stripping from generating mathematical artifacts.

[0067] If the current normalized energy percentage is determined to be higher than 5%, and the value of the iteration counter has not reached the preset maximum threshold (in this embodiment, the maximum number of iterations is preferably set to 10. This maximum threshold can be adaptively adjusted according to the sampling frequency and power grid monitoring accuracy requirements, aiming to ensure that the residual space is locked in time for weak feature extraction after the dominant energy component is basically cleared), then it is determined that the current feature stripping is insufficient, and a recursive call instruction is initiated. Before executing the program jump, the microprocessor performs an environment initialization operation, that is, clears the intermediate register used to store the internal integral distribution vector and resets the global maximum value search pointer to reduce the interference of the previous round's matching result on the current round. Subsequently, through the address jump of the instruction register, the data flow is guided back to the aforementioned projection operator operation step. This dual convergence criterion design ensures that even when encountering nonlinear distortion conditions not covered by the atomic library, it can be forcibly exited within a limited time to avoid falling into an infinite loop.

[0068] If any of the above termination conditions are met (energy drops below 5% or 10 iterations are reached), the final residual sequence is locked, and its memory address is passed to the subsequent feature detection operator to generate a residual impulse feature vector reflecting weak non-stationary features.

[0069] The present invention aims to achieve adaptive layering and stripping of signal components through a dual feedback mechanism of "energy-number".

[0070] After each round of subtraction stripping, the residual sequence in the buffer is read and a logical judgment is performed. When the normalized energy ratio is higher than 5% and the iteration count is less than 10, the microprocessor clears the intermediate register and resets the search pointer. The residual sequence is mapped to the next round of input and returned to the projection operation through instruction jump. If any termination condition is met, the final residual is locked and passed to the feature detection operator.

[0071] By using dual-criteria collaborative control, while ensuring that complex disturbances are fully deconstructed, it can effectively avoid the problems of algorithm non-convergence or dead loops that may occur under nonlinear distortion conditions, thereby improving the logical robustness and computational determinism when running on the embedded monitoring terminal.

[0072] Further, until the energy modulus of the residual sequence satisfies the preset convergence criterion, the following steps are taken: calculating the square of the L2 norm of the residual sequence output in the current iteration to obtain the current energy modulus; calculating the ratio of the current energy modulus to the total energy modulus of the multi-source power grid sampling sequence to output the normalized energy ratio; performing a comparison operation between the normalized energy ratio and a preset calculation accuracy limit; if the normalized energy ratio is lower than the calculation accuracy limit, then it is determined that the convergence criterion is satisfied.

[0073] Specifically, after outputting the residual sequence in each iteration, the digital signal microprocessor first calculates the square of the L2 norm of the sequence. This involves traversing the 512 sampling points in the buffer and performing the squaring operation point by point. To address the potential risk of accumulation overflow in the 32-bit register of the embedded microprocessor, this embodiment employs the following strategy in the squaring accumulation logic: using a 64-bit wide accumulation register for sum storage, or pre-scaling the sampling point values ​​before performing the squaring operation (e.g., shifting the entire value 4 bits to the right), to ensure that the accumulated sum remains within the valid numerical representation range even when dealing with large-amplitude signals output by the 16-bit analog-to-digital converter.

[0074] Subsequently, the ratio of the current energy magnitude to the total energy magnitude of the multi-source power grid sampling sequence (i.e., the L2 norm of the original signal) is calculated, and the normalized energy proportion is output. This proportion directly reflects the percentage of remaining energy in the current residual relative to the total energy of the original signal.

[0075] Finally, the normalized energy percentage is compared with a preset calculation accuracy limit. In this embodiment, the calculation accuracy limit is set to 0.05 (i.e., 5%). Analyzing the physical energy distribution of the power quality signal, if the residual energy of the dominant disturbance components (such as sags and harmonics) drops below 5% after recursive stripping, it means that the main deterministic features of the signal have been extracted, and the second-order statistical characteristics of the residual sequence tend towards Gaussian white noise. Anchoring the threshold at 0.05 avoids the inclusion of incompletely stripped dominant components in the residual due to an excessively high threshold, and also prevents overfitting at the noise level due to an excessively low threshold (e.g., <0.01), resulting in mathematical artifacts. If the normalized energy percentage is below 0.05, the current residual energy is determined to have shrunk to the noise level or a very weak level, satisfying the convergence criterion, and the iteration is terminated.

[0076] This invention provides a scientific convergence criterion for recursive iteration by quantifying the normalized energy proportion. In the calculation of the square of the L2 norm, a 64-bit accumulation or pre-scaling strategy is employed, effectively reducing the risk of register overflow under large-amplitude signals and ensuring the stability of numerical operations. This method uses the normalized proportion to intuitively reflect the decay state of the residual energy. Combined with a preset calculation accuracy limit, the algorithm can exit the iteration in a timely manner when the energy drops to the noise level. While ensuring the depth of the layered deconstruction of complex perturbations, it effectively balances calculation accuracy and the computational efficiency of the embedded processor.

[0077] Further, detecting and outputting the residual pulse feature vector from the residual sequence includes: using a high-pass filter operator to filter out low-frequency residual components in the residual sequence and outputting a high-frequency component sequence; performing time-domain segmentation mapping on the high-frequency component sequence to extract a set of sampling points whose amplitude jump rate exceeds a preset rate of change benchmark; using a pulse positioning operator to determine the distribution range of the sampling point set on the time axis and outputting a pulse position index; and based on the pulse position index, extracting the corresponding numerical components from the residual sequence to construct and generate the residual pulse feature vector.

[0078] Specifically, a high-pass filter operator is used to filter out low-frequency residual components in the residual sequence. Since steady-state harmonics in power systems are typically distributed below 2kHz (i.e., within the 40th harmonic), the digital signal microprocessor invokes a preset 32nd-order finite impulse response (FIR) high-pass filter. In this embodiment, with a sampling frequency of 12.8kHz, the cutoff frequency of the FIR high-pass filter is set to 2000Hz. Because the aforementioned iterative process has already removed the main power frequency and its low-order harmonic components, this step aims to reduce the potentially slow fluctuation baseline in the residual, outputting a high-frequency component sequence containing information about rapid transients. Simultaneously, the parameters of the high-pass filter can be adaptively adjusted according to changes in the actual sampling frequency, based on the Nyquist theorem, to maintain the same physical cutoff bandwidth.

[0079] Subsequently, time-domain segmentation mapping is performed on the high-frequency component sequence, using a sliding window with a length of 8 sampling points. The sum of the absolute differences between adjacent sampling points within the window is calculated point by point. To ensure the consistency of the dimensionality of the judgment criteria, this embodiment sets the quantization discrimination criterion as follows: when the sum of the absolute differences between the 8 sampling points within the window is greater than 5% of the nominal amplitude, the window is determined to cover potential pulse characteristics.

[0080] Next, the pulse position index is determined using a pulse positioning operator. This embodiment employs one-dimensional connectivity judgment logic to perform cluster analysis: transition points with adjacent intervals of less than four sampling points on the time axis are grouped into the same pulse interval. This effectively avoids misclassifying a single pulse as multiple discrete points while ensuring real-time performance.

[0081] In particular, in order to compensate for the phase lag generated by the filter operator, group delay compensation is performed when indexing the output pulse position: according to the characteristics of the 32nd order finite impulse response high-pass filter, the initially located interval index is uniformly shifted forward by 16 sampling points (i.e., half the order), so that the located high-frequency change position and the original residual position are physically aligned on the time axis.

[0082] Finally, based on the compensated pulse position index, the corresponding numerical components are truncated from the residual sequence. To meet the standardization requirements of the input data dimension for subsequent classification algorithms, this embodiment sets the fixed length of the residual pulse feature vector to 64 dimensions: if the length of the truncated numerical component is less than 64 bits, zero vector padding is performed at the end of the sequence; if it exceeds 64 bits, 32 bits are truncated on both sides of the pulse peak to construct the residual pulse feature vector.

[0083] The present invention aims to refine the extraction of high-frequency pulse features from residual sequences through a "filter positioning-original value truncation" mechanism with time delay compensation.

[0084] First, a 32nd-order high-pass filter is used to extract high-frequency components from the residual. An 8-sampling-point sliding window is used to detect the amplitude jump rate to identify potential pulse points. Then, pulse intervals are generated by one-dimensional connectivity determination logic clustering. Group delay translation compensation of 16 sampling points is performed according to the filter characteristics. Finally, the original residual is truncated and standardized according to the corrected index and reconstructed into a 64-dimensional feature vector.

[0085] It cleverly solves the phase lag problem caused by digital filtering. Through the separation strategy of "positioning and truncation", it achieves precise alignment between the high-frequency mutation position and the physical characteristics of the original residual. It not only preserves the transient details of the disturbance, but also provides a unified format through standardized reconstruction, which improves the reliability of subsequent spatial coherence analysis and composite disturbance identification.

[0086] Furthermore, referring to Figure 3 The process of performing the coherence calculation and removing pseudo-features generated by random noise includes: obtaining the time-frequency atom distributions of each distributed node physically adjacent to the current node within a preset propagation delay compensation window; performing a sliding window search on the residual pulse feature vector using a cross-correlation operator, and performing a dot product operation with each of the time-frequency atom distributions respectively, outputting a spatial coherence score sequence; labeling the values ​​in the spatial coherence score sequence that are below a preset consistency threshold as pseudo-feature components, and labeling the remaining values ​​that meet the preset consistency threshold as valid perturbation components.

[0087] Specifically, the time-frequency atom distribution of each physically adjacent distributed node within a preset propagation delay compensation window is obtained. In this embodiment, each distributed node is equipped with a timing module based on the BeiDou satellite navigation system to ensure that the synchronization sampling time deviation of all nodes in the network is less than 50 microseconds. Under the premise of this high-precision clock synchronization, considering that the propagation speed of electromagnetic waves in the power distribution feeder is about two-thirds of the speed of light, combined with the microsecond-level propagation delay caused by the single-side power supply radius of a 10kV power distribution network typically not exceeding 10 kilometers, and network communication jitter, 2 milliseconds is sufficient to cover the time difference of signal propagation across nodes. Therefore, the current node retrieves the monitoring data of the adjacent nodes within a 2-millisecond window before and after the corresponding timestamp; the adjacent nodes also regularize the residual pulse feature vectors they detect into 64-dimensional discrete numerical vectors for interaction.

[0088] Subsequently, a sliding window search is performed on the residual pulse feature vector using a cross-correlation operator. The digital signal microprocessor performs a normalized inner product operation on the local residual pulse feature vector and the feature waveforms transmitted from adjacent nodes, outputting a spatial coherence score. The calculation logic of the normalized inner product is as follows: calculate the absolute value of the dot product of the two vectors and divide it by the product of the magnitudes of the two vectors. Through this magnitude normalization process, the range of the spatial coherence score is strictly limited to between 0 and 1. This processing method reduces amplitude fluctuation interference between different nodes caused by differences in sensor ratios or transmission attenuation, making the score depend only on the similarity of waveform morphology. If the waveform morphology of two locations is highly consistent, the score approaches 1.0.

[0089] Finally, values ​​in the spatial coherence score sequence below a preset consistency threshold (0.65) are designated as pseudo-feature components. Since random noise or electromagnetic interference generated by local hardware is highly independent in space (i.e., it does not have cross-line propagation characteristics), its normalized inner product score between adjacent nodes is usually much lower than 0.65; these components are zeroed out. Values ​​with scores higher than 0.65 are designated as valid perturbation components, retained, and included in subsequent synthesis processes.

[0090] The present invention aims to use the spatial coherence verification of distributed nodes to achieve the elimination of false features and the preservation of true features.

[0091] First, the time-frequency atomic distribution of the current node and its physical neighbors within a 2-millisecond delay compensation window is obtained through the BeiDou timing module. Then, the local residual pulse feature vector and the feature waveforms of the neighboring nodes are subjected to normalized inner product operation to output the spatial coherence score sequence. Components with scores lower than the 0.65 consistency threshold are labeled as pseudo-feature components and subjected to zeroing processing, while only the effective perturbation components that meet the threshold are retained.

[0092] By aligning features in the spatial dimension and utilizing the difference between the cross-line propagation characteristics of real disturbances and the spatial independence of random noise, false signals generated by sensor jitter or local environmental degradation in single-point monitoring can be effectively eliminated, retaining only the real power grid disturbance components with propagation characteristics. This improves the robustness of distributed power quality monitoring terminals in identifying composite disturbances and the overall anti-interference capability of the monitoring network.

[0093] Further, the process of reconstructing and generating the composite perturbation feature matrix includes: extracting the amplitude envelope parameters and center frequency parameters of the main conductive energy component output in each iteration to construct a background feature set; mapping the effective perturbation component to the coordinate space where the background feature set is located according to the sampling timestamp index; calculating the coupling weight between the effective perturbation component and the background feature set; performing parameter alignment and nonlinear recombination processing to generate a cooperative feature vector; mapping the cooperative feature vector to a matrix operator template according to a preset topological dimension, wherein, based on the difference between the actual number of adjacent nodes and the preset column dimension, adaptive filling or filtering operations are performed on the cooperative feature columns in the matrix operator template to output the composite perturbation feature matrix.

[0094] Specifically, the amplitude envelope parameters and center frequency parameters of the main conductive energy component output from each iteration are extracted to construct a background feature set. The digital signal microprocessor extracts scalar values ​​reflecting the voltage sag depth, harmonic component amplitude, and oscillation center frequency from the atomic parameters successfully extracted in each iteration. For example, if a voltage sag component is extracted in the first iteration, its corresponding amplitude reduction ratio (e.g., 0.7 times the nominal value) and duration parameter are stored in the background feature vector. These parameters constitute the macroscopic skeleton of the signal.

[0095] Subsequently, the effective perturbation components are mapped to the coordinate space of the background feature set according to the sampling timestamp index. During this process, the digital signal microprocessor substitutes the scalar parameters into the corresponding elementary analytical expressions based on the scalar values ​​in the background feature set, and parsely generates the corresponding continuous-time elementary function sequence (such as a standard sine function or an exponentially decaying function), thereby obtaining the analytical solution of the background components at any given time. Since group delay compensation has been performed in the preceding steps, the detected 64-dimensional effective perturbation components are located on the time axis of the background features.

[0096] By calculating the coupling weights between the effective perturbation components and the background feature set, parameter alignment and nonlinear recombination are performed. In this embodiment, the coupling weights are defined as the normalized instantaneous voltage change rate (i.e., the first derivative) of the background component sequence generated by analysis at the pulse entry moment. The nonlinear recombination uses a cross-product operator, that is, by calculating the product of the 64-dimensional effective perturbation components and the corresponding time-time coupling weights, a cross-feature term reflecting the perturbation impact intensity is generated. The composite perturbation features are not a linear superposition of the components. Through cross-product, the asymmetric distortion characteristics (i.e., modulation effect) of the pulse under different background phases can be effectively captured, the intensity of superposition of weak pulses under different perturbation backgrounds can be quantified, and the dynamic coupling relationship between multi-source perturbations can be captured.

[0097] Finally, the collaborative feature vectors are mapped to the matrix operator template according to the preset topological dimension. In this embodiment, a fixed dimension is constructed. The composite disturbance feature matrix is ​​constructed by setting 8 rows of physical parameters to build a multi-dimensional fingerprint covering the entire life cycle of the disturbance; the 16 columns are set to cover most typical distribution network topologies (usually with no more than 6 branches) while also taking into account the depth of recursive decomposition. The row dimension... These correspond to eight types of physical parameters: normalized amplitude, center frequency, phase offset, envelope attenuation factor, pulse transition rate, spatial coherence score, waveform skewness, and kurtosis; column dimensions. This is composed of 10 columns of iterative component features and 6 columns of neighboring node collaborative features. During the mapping process, the 8 types of physical parameters (such as amplitude, center frequency, etc.) contained in the collaborative feature vector are sequentially filled into the corresponding rows of the matrix. To address the dynamic differences in the distributed topology, adaptive alignment of the neighboring node columns is performed: Let the actual number of neighboring nodes be... If the actual number of adjacent nodes Then for the remaining elements in the matrix Column collaborative features perform zero-filling; if Then, based on the spatial coherence score, the features of the first 6 neighboring nodes are selected from high to low and loaded. If the iteration terminates prematurely, the subsequent corresponding columns are zeroed out to maintain the constant matrix dimension.

[0098] The output generates the composite disturbance feature matrix. This matrix, serving as the final classification input, not only contains the independent features of a single disturbance but also captures the nonlinear coupling relationship between multiple disturbances through "cooperative" logic. By inputting this matrix into subsequent classification operators, the monitoring terminal can achieve extremely high-precision identification of complex composite conditions such as "voltage dips superimposed with transient oscillations," solving the technical problem of feature ambiguity in traditional methods when processing non-stationary, multi-source superimposed signals.

[0099] This invention, by constructing a complete atomic library and combining it with a recursive stripping algorithm, achieves hierarchical deconstruction of composite power quality disturbance components, effectively solving the technical challenge of extracting weak disturbance features under strong background energy masking. By introducing a spatial coherence verification mechanism using distributed nodes, the algorithm achieves the removal of false data from disturbance components, improving the fidelity of feature extraction. The final reconstructed composite disturbance feature matrix can characterize the nonlinear coupling relationship between disturbances in multiple dimensions, providing structured and highly discriminative data support for high-precision modeling and identification of complex composite conditions in heterogeneous data environments, and enhancing the robustness of monitoring terminals in complex power grid environments.

[0100] Example 2:

[0101] In this embodiment, the aforementioned power quality processing method is deployed in a distributed monitoring terminal of a 10kV distribution network intelligent transformation line. This line includes high-capacity inductive loads (such as elevator drive motors) and exposed outdoor overhead line sections.

[0102] During an actual monitoring operation, the distributed monitoring terminal acquired a composite perturbation sequence containing 512 sampling points. First, the microprocessor initiated an atom library matching process, performing projection onto an overcomplete atom library discretized with a 0.05 logarithmic step size. In the first iteration, the algorithm identified a decaying sinusoidal atom with a center frequency of 50Hz and a significantly decreasing amplitude, identifying it as the dominant conductive energy component. This physically corresponds to the voltage dip in the power grid caused by the instantaneous startup of a large-capacity inductive load.

[0103] Subsequently, a recursive subtraction stripping logic is executed to remove the aforementioned concave components from the original sequence. In the residual sequence, the weak transient features that were originally masked by amplitude fluctuations are clearly highlighted. The microprocessor calls a 32nd-order finite impulse response high-pass filter to extract components above 2000Hz from the residual and compensates for the group delay of 16 sampling points, locating a high-frequency pulse with a duration of approximately several hundred microseconds at the corresponding point on the time axis. This pulse, after spatial coherence verification after synchronization with BeiDou, is compared with the waveforms transmitted back from adjacent nodes to obtain a normalized inner product score of 0.82, which is higher than the consistency threshold of 0.65. It is determined to be a genuine induced lightning electromagnetic pulse, thus identifying and preserving this effective disturbance component.

[0104] Next, using a cross-product operator, the peak energy of the pulse is nonlinearly recombined with the instantaneous voltage change rate at the beginning of the voltage dip, generating a cooperative feature vector that captures the dynamic coupling weights generated by the pulse at the moment of waveform drop. Finally, a... A composite perturbation feature matrix of dimension 1, wherein specific columns of the matrix record the envelope parameters, pulse position indexes and spatial coherence scores stripped out in that round.

[0105] This invention can separate the extracted dominant component from the residual pulse component, and the generated composite disturbance feature matrix provides highly discriminative structured data for subsequent classification and identification, realizing the restoration of composite operating conditions and meeting the technical requirements of intelligent terminals of distribution networks for real-time and efficient detection of composite disturbances.

[0106] The structures and processes shown in this invention are for illustrative purposes only. Those skilled in the art should understand that any non-substantial modifications, substitutions, or structural adjustments made to the relevant parameters, modules, or logical topologies without departing from the essential innovative spirit of this invention shall not affect their inclusion within the scope of the claims and their equivalents.

Claims

1. A collaborative modeling method for power quality composite disturbances based on multi-source data, characterized in that, include: Based on the attenuation and oscillation time-frequency characteristics of multi-source power grid sampling sequences collected by distributed nodes, an overcomplete atomic library containing a set of primitive functions is constructed. The multi-source power grid sampling sequence is projected onto the overcomplete atom library to perform projection operator operations. By calculating the inner product of each atom in the overcomplete atom library with the multi-source power grid sampling sequence, the main conductive energy component with the highest energy contribution is identified and output. The main conductive power component is recursively subtracted from the multi-source power grid sampling sequence to output a residual sequence; The residual sequence is used as the input for the next iteration. The projection operator is returned to perform cyclic stripping until the energy magnitude of the residual sequence meets the preset convergence criterion. The residual pulse feature vector is then detected and output from the residual sequence. The residual pulse feature vector is combined with the time-frequency atomic distribution of adjacent nodes in the distributed nodes, and coherence calculation is performed. By verifying the redundancy of the residual pulse feature vector in the spatial dimension, false features generated by random noise are eliminated, and a composite perturbation feature matrix is ​​reconstructed. The convergence process continues until the energy modulus of the residual sequence satisfies a preset convergence criterion, including: calculating the square of the L2 norm of the residual sequence output in the current iteration to obtain the current energy modulus; calculating the ratio of the current energy modulus to the total energy modulus of the multi-source power grid sampling sequence to output the normalized energy ratio; comparing the normalized energy ratio with a preset calculation accuracy limit, and if the normalized energy ratio is lower than the calculation accuracy limit, then the convergence criterion is satisfied.

2. The method for collaborative modeling of power quality composite disturbances based on multi-source data according to claim 1, characterized in that, The process of constructing an overcomplete atomic library based on the distributed nodes includes: performing analog-to-digital conversion on the original power signal using the distributed nodes deployed at the power distribution feeder end, and outputting the multi-source power grid sampling sequence; determining a disturbance physical parameter space including the envelope attenuation coefficient range and the oscillation center frequency range based on the statistical characteristics of the multi-source power grid sampling sequence; injecting the attenuation factor distribution, frequency operator distribution, and displacement factor distribution generated by mapping the disturbance physical parameter space as parameters to be optimized into a preset primitive function set; performing multi-scale discretization iterative sampling on the primitive function set to generate an atom set covering different time-frequency granularities, and constructing the overcomplete atomic library.

3. The method for collaborative modeling of power quality composite disturbances based on multi-source data according to claim 1, characterized in that, Projecting the multi-source power grid sampling sequence onto the overcomplete atom library and performing projection operator operations includes: performing conjugate transpose multiplication on the multi-source power grid sampling sequence and each atom in the atom set of the overcomplete atom library to output a sequence of inner product real numbers reflecting the degree of waveform matching; performing absolute value processing on the inner product real number sequence to generate an inner product modulus distribution vector; searching for the global maximum value in the inner product modulus distribution vector, and labeling the index value corresponding to the global maximum value as the optimal atom index for the current iteration.

4. The method for collaborative modeling of power quality composite disturbances based on multi-source data according to claim 3, characterized in that, Identifying and outputting the main conductive energy component with the highest energy contribution includes: retrieving the corresponding target primitive waveform function from the overcomplete atom library according to the optimal atom index; performing a scalar multiplication operation between the target primitive waveform function and the weight coefficient at the corresponding index position in the inner product real number sequence to generate the main conductive energy component.

5. The method for collaborative modeling of power quality composite disturbances based on multi-source data according to claim 1, characterized in that, The process of performing the recursive subtraction stripping and outputting the residual sequence includes: taking the multi-source power grid sampling sequence as the signal sequence to be processed in the current round, and taking the main conductive energy component as the deduction item in the current round; performing point-by-point numerical subtraction operation on the signal sequence to be processed and the deduction item using the subtraction operator, and outputting the temporary residual vector of the current round; and labeling the temporary residual vector as the residual sequence.

6. The method for collaborative modeling of power quality composite disturbances based on multi-source data according to claim 1, characterized in that, The residual sequence is used as the input for the next iteration to perform loop stripping, including: determining whether the energy magnitude of the residual sequence is higher than the preset convergence criterion; if the determination result is higher than the preset convergence criterion, a recursive call instruction is initiated to map the residual sequence output in the current round to the signal sequence to be processed in the next iteration loop, and the projection operator operation is returned to be executed; if the determination result is not higher than the preset convergence criterion, the iteration loop is terminated and the final residual sequence is locked, and the final residual sequence is input to the feature detection operator to generate the residual pulse feature vector.

7. The method for collaborative modeling of power quality composite disturbances based on multi-source data according to claim 1, characterized in that, Detecting and outputting a residual pulse feature vector from the residual sequence includes: using a high-pass filter operator to filter out low-frequency residual components in the residual sequence and outputting a high-frequency component sequence; performing time-domain segmentation mapping on the high-frequency component sequence to extract a set of sampling points whose amplitude jump rate exceeds a preset rate of change benchmark; using a pulse positioning operator to determine the distribution range of the sampling point set on the time axis and outputting a pulse position index; and based on the pulse position index, extracting the corresponding numerical components from the residual sequence to construct and generate the residual pulse feature vector.

8. The method for collaborative modeling of power quality composite disturbances based on multi-source data according to claim 1, characterized in that, The process of performing the coherence calculation and removing pseudo-features generated by random noise includes: obtaining the time-frequency atom distributions of each distributed node physically adjacent to the current node within a preset propagation delay compensation window; performing a sliding window search on the residual pulse feature vector using a cross-correlation operator, and performing a dot product operation with each of the time-frequency atom distributions respectively to output a spatial coherence score sequence; labeling the values ​​in the spatial coherence score sequence that are below a preset consistency threshold as pseudo-feature components, and labeling the remaining values ​​that meet the preset consistency threshold as valid perturbation components.

9. The method for collaborative modeling of power quality composite disturbances based on multi-source data according to claim 8, characterized in that, The process of reconstructing and generating the composite perturbation feature matrix includes: extracting the amplitude envelope parameters and center frequency parameters of the main conductive energy component output in each iteration to construct a background feature set; mapping the effective perturbation component to the coordinate space where the background feature set is located according to the sampling timestamp index; calculating the coupling weight between the effective perturbation component and the background feature set; performing parameter alignment and nonlinear recombination processing to generate a cooperative feature vector; and mapping the cooperative feature vector to a matrix operator template according to a preset topological dimension, wherein, based on the difference between the actual number of adjacent nodes and the preset column dimension, adaptive filling or filtering operations are performed on the cooperative feature columns in the matrix operator template to output the composite perturbation feature matrix.

Images