Electric energy quality comprehensive monitoring method and system based on real-time waveform analysis

Abstract

The invention discloses an electric energy quality comprehensive monitoring method and system based on real-time waveform analysis. The method comprises the following steps: S1, acquiring an analog voltage signal acquired by a mutual inductor; s2, preprocessing to obtain a voltage waveform data segment; s3, performing multi-dimensional feature extraction to obtain a voltage waveform multi-dimensional feature vector; s4, extracting a voltage RMS effective value, a total harmonic distortion rate and wavelet packet energy; s5, inputting the voltage RMS effective value, the total harmonic distortion rate and the wavelet packet energy into a grading event pre-detection model to obtain an event feature subset coding vector; and S6, enabling the event feature subset coding vector to pass through an intelligent event fine classification model to obtain an electric energy event accurate classification result. According to the electric energy quality comprehensive monitoring method and system based on real-time waveform analysis disclosed by the invention, the accuracy and efficiency of electric energy quality event monitoring are greatly improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of power quality monitoring, and particularly relates to a power quality comprehensive monitoring method and system based on real-time waveform analysis. BACKGROUND

[0002] With the increasing dependence on electricity in modern industry, commerce and daily life, power quality has become a key indicator of the performance of the power supply system. High-quality power is the basis for ensuring the efficient, safe and stable operation of various electrical equipment. However, the power system is inevitably disturbed by various internal and external factors during operation, such as the nonlinearity and suddenness of the load, the large number of power electronic devices connected, and natural phenomena, etc. These factors can cause a series of power quality problems such as voltage sag, voltage swell, interruption, harmonics, three-phase imbalance, voltage fluctuation and flicker, and transient overvoltage. These problems not only affect the normal operation of electrical equipment, reduce production efficiency, and shorten equipment life, but also may cause equipment damage, data loss, system failure and other serious consequences, resulting in huge economic losses. Therefore, in order to ensure the reliability of the power system and the safety of electricity, it is crucial to build a comprehensive monitoring scheme that can comprehensively, real-time and accurately monitor the power quality status, which can provide key data support for early warning, fault diagnosis, cause analysis and treatment measures of power quality problems.

[0003] In existing power quality monitoring practices, some traditional methods mainly focus on threshold monitoring of a few key power quality parameters. For example, some monitoring devices may only calculate the root mean square (RMS) value of the voltage to determine whether voltage sag or swell occurs, or only use fast Fourier transform (FFT) analysis to assess whether the total harmonic distortion (THD) is excessive. Although this method can reflect some specific power quality problems to some extent, its defects are also obvious: first, it usually cannot capture complex and composite power quality events, such as voltage sag accompanied by severe harmonic distortion, or the superposition of transient oscillation and voltage fluctuation; second, for some transient disturbances with extremely short duration but strong destructive power, such as sharp pulses or oscillation transients, this method based on steady-state parameter analysis often has difficulty in effective detection and characterization; third, its analysis dimension is single, lacking in-depth mining of waveform detail features, resulting in insufficient ability to accurately classify event types and difficulty in providing refined diagnostic information.

[0004] Another common technical solution is event-triggered recording analysis, for example, when the voltage effective value exceeds the preset range, the system records a segment of waveform data for subsequent offline analysis. Although this kind of solution can obtain the original waveform at the time of event occurrence, its triggering mechanism is often simple and may miss some hidden disturbances that do not meet the preset triggering conditions. At the same time, the subsequent analysis often relies on manual experience or relatively simple signal processing algorithms, making it difficult to achieve automatic, fast and accurate classification of various power quality events, especially when facing massive monitoring data and complex disturbance types, its efficiency and accuracy are difficult to meet the needs of modern power grid fine management, not to mention comprehensive evaluation and intelligent identification of various disturbance characteristics.

[0005] Therefore, in view of the above problems, further improvement is needed. SUMMARY

[0006] The main purpose of the present application is to provide a power quality comprehensive monitoring method and system based on real-time waveform analysis, which first extracts a comprehensive voltage waveform multi-dimensional feature vector from real-time voltage waveform to completely capture power quality disturbance information. Subsequently, a hierarchical processing strategy is innovatively adopted: through a hierarchical event pre-detection model, key features such as voltage RMS effective value, total harmonic distortion rate and wavelet packet energy are used to quickly screen out suspected events and generate event feature subset encoding vectors; then the event feature subset encoding vectors are sent to an intelligent event fine classification model for in-depth analysis, thereby realizing accurate analysis and classification of power events. In this way, by extracting multi-dimensional features from real-time waveform and combining the two-stage model of hierarchical pre-detection and intelligent fine classification, the present application greatly improves the accuracy and efficiency of power quality event monitoring. It can not only accurately identify single events such as voltage sag and harmonic overlimiting, but more importantly, through intelligent correlation analysis of event features and context features, it can effectively distinguish and diagnose complex and transient power quality disturbances that are difficult to handle by traditional methods, achieving comprehensive monitoring and fine diagnosis, reducing dependence on manual experience, and improving the automation and intelligence level of the monitoring system.

[0007] To achieve the above purpose, the present application provides a power quality comprehensive monitoring method based on real-time waveform analysis, comprising the following steps: Step S1: acquiring an analog voltage signal collected by a mutual inductor; Step S2: preprocessing the analog voltage signal to obtain a voltage waveform data segment; Step S3: multi-dimensional feature extraction is performed on the voltage waveform data segment to obtain a voltage waveform multi-dimensional feature vector; Step S4: extracting voltage RMS effective value, total harmonic distortion rate and wavelet packet energy from the voltage waveform multi-dimensional feature vector; Step S5: inputting the voltage RMS effective value, total harmonic distortion and wavelet packet energy into a hierarchical event pre-detection model to obtain an event feature subset coding vector; Step S6: inputting the event feature subset coding vector into an intelligent event fine classification model to obtain an electric energy event accurate classification result.

[0008] As a further preferred technical solution of the above technical solution, in step S2, the analog voltage signal is sampled at a predetermined sampling frequency to obtain a plurality of digital voltage sampling points; the plurality of digital voltage sampling points are low-pass filtered to obtain the voltage waveform data segment.

[0009] As a further preferred technical solution of the above technical solution, step S3 is specifically implemented as the following steps: Step S31: calculating the voltage RMS effective value of the voltage waveform data segment; Step S32: performing FFT analysis on the voltage waveform data segment to obtain the fundamental amplitude, third harmonic amplitude, fifth harmonic amplitude and total harmonic distortion; Step S33: extracting the instantaneous feature from the voltage waveform data segment to obtain the wavelet packet energy; Step S34: extracting the voltage flicker degree and voltage unbalance degree from the voltage waveform data segment; Step S35: arranging the voltage RMS effective value, fundamental amplitude, third harmonic amplitude, fifth harmonic amplitude, total harmonic distortion, wavelet packet energy, voltage flicker degree and voltage unbalance degree into the voltage waveform multi-dimensional feature vector.

[0010] As a further preferred technical solution of the above technical solution, step S5 is specifically implemented as the following steps: Step S51: judging whether the voltage RMS effective value is less than a preset threshold, if the condition is met, triggering a suspected transient event rule to include the voltage RMS effective value in the event feature subset coding vector; Step S52: judging whether the total harmonic distortion is greater than a preset threshold, if the condition is met, triggering a suspected harmonic event rule to include the total harmonic distortion in the event feature subset coding vector; Step S53: judging whether the wavelet packet energy is greater than a preset threshold, if the condition is met, triggering a suspected transient event rule to include the wavelet packet energy in the event feature subset coding vector.

[0011] As a further preferred technical solution of the above technical solution, step S6 is specifically implemented as the following steps: Step S61: excluding the event feature subset coding vector from the voltage waveform multi-dimensional feature vector to obtain a context feature vector; Step S62: based on the context feature vector, event feature saliency is performed on the event feature subset coding vector to obtain a saliency event feature subset coding vector; Step S63: the saliency event feature subset coding vector is input into a pre-trained Random Forest classification model to obtain an accurate classification result of the electric energy event.

[0012] As a further preferred technical solution of the above technical solution, step S62 is specifically implemented as the following steps: Step S621: linear transformation is performed on the event feature subset coding vector to obtain a transformed event feature subset coding vector; Step S622: an event feature context global fine-grained interaction feature matrix of the transformed event feature subset coding vector relative to the context feature vector is calculated; Step S623: the event feature context global fine-grained interaction feature matrix is decomposed by column vector to obtain a sequence distribution of event feature context local response coding vectors; Step S624: the sequence distribution of the event feature context local response coding vectors is input into a forward LSTM model-based inference sequence derivation module to obtain the saliency event feature subset coding vector.

[0013] As a further preferred technical solution of the above technical solution, step S622 is specifically implemented as the following steps: Step S6221: the transformed event feature subset coding vector and the event feature subset coding vector are multiplied by transposition to obtain an event feature context correlation matrix; Step S6222: a dynamic weight matrix and a projection matrix are determined; after element-level multiplication of the event feature context correlation matrix and the dynamic weight matrix, an event feature context asymmetric interaction representation matrix is obtained through a tanh function; Step S6223: the projection matrix and the event feature context correlation matrix are input into a Sigmoid function for activation, and then multiplied to obtain an event feature context gating matrix; Step S6224: the event feature context asymmetric interaction representation matrix and the event feature context gating matrix are fused to obtain an event feature context global fine-grained interaction feature matrix.

[0014] As a further preferred technical solution of the above technical solution, step S6224 is specifically implemented as: sequentially projecting consistency verification on the event feature context asymmetric interaction representation matrix and the event feature context gating matrix to obtain a transition matrix; performing distribution state degeneration operation on the transition matrix to obtain a correction matrix; and based on the correction matrix, correcting the event feature context asymmetric interaction representation matrix and the event feature context gating matrix respectively, and then performing position-wise addition to obtain an event feature context global fine-grained interaction feature matrix.

[0015] To achieve the above object, the application further provides an electric energy quality comprehensive monitoring system based on real-time waveform analysis, comprising: an analog voltage signal acquisition module, configured to acquire analog voltage signals collected by a mutual inductor; An analog voltage signal preprocessing module is configured to preprocess the analog voltage signals to obtain voltage waveform data segments. A multi-dimensional feature extraction module is configured to extract multi-dimensional features from the voltage waveform data segments to obtain voltage waveform multi-dimensional feature vectors. A voltage waveform multi-dimensional feature analysis module is configured to extract voltage RMS effective value, total harmonic distortion rate and wavelet packet energy from the voltage waveform multi-dimensional feature vectors. An event feature pre-detection and coding module is configured to input the voltage RMS effective value, total harmonic distortion rate and wavelet packet energy into a hierarchical event pre-detection model to obtain an event feature subset coding vector. An event feature coding classification analysis module is configured to pass the event feature subset coding vector through an intelligent event fine classification model to obtain an electric energy event accurate classification result.

[0016] The application has the following beneficial effects: Firstly, a comprehensive voltage waveform multi-dimensional feature vector is extracted from the real-time voltage waveform to fully capture the disturbance information of power quality. Subsequently, a hierarchical processing strategy is innovatively adopted: through a hierarchical event pre-detection model, key features such as voltage RMS effective value, total harmonic distortion rate and wavelet packet energy are used to quickly screen out suspected events and generate event feature subset encoding vectors; then the event feature subset encoding vectors are sent to an intelligent event fine classification model for in-depth analysis of fine and situational, so as to realize accurate analysis and classification of power events. In this way, through multi-dimensional feature extraction of real-time waveform and combined with the two-stage model of hierarchical pre-detection and intelligent fine classification, the accuracy and efficiency of power quality event monitoring are greatly improved. It can not only accurately identify single events such as voltage sag and harmonic overlimiting, but more importantly, through intelligent correlation analysis of event features and context features, it can effectively distinguish and diagnose composite and transient power quality disturbances that are difficult to handle by traditional methods, realize the comprehensiveness of monitoring and the refinement of diagnosis, reduce the dependence on artificial experience, and improve the automation and intelligence level of the monitoring system. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a schematic flowchart of the method of the present invention.

[0018] Figure 2 is a schematic diagram of data flow of the present invention.

[0019] Figure 3 is a schematic flowchart of step S3 of the present invention.

[0020] Figure 4 is a schematic flowchart of step S5 of the present invention.

[0021] Figure 5 is a schematic flowchart of step S6 of the present invention.

[0022] Figure 6 is a schematic block diagram of the system of the present invention.

[0023] The reference signs include: 110, analog voltage signal acquisition module; 120, analog voltage signal preprocessing module; 130, multi-dimensional feature extraction module; 140, voltage waveform multi-dimensional feature analysis module; 150, event feature pre-detection and coding module; 160, event feature coding classification analysis module. DETAILED DESCRIPTION

[0024] The following description is presented to enable any person skilled in the art to practice the application as claimed. The preferred embodiments disclosed herein are only examples of the many possible variations of the present application. Other variations of the present application are possible and will be apparent to those skilled in the art upon reading the following description. The basic principles of the present application defined in the following description are applicable to other embodiments, variations, modifications, equivalents, and other implementations of the present application without departing from the spirit and scope of the present application.

[0025] In the preferred embodiments of the present application, it should be noted by those skilled in the art that the mutual inductor and the like involved in the present application can be regarded as prior art.

[0026] Preferred embodiments.

[0027] As Figures 1-5 shown, the present application discloses a power quality comprehensive monitoring method based on real-time waveform analysis, comprising the following steps: Step S1: acquiring an analog voltage signal collected by a mutual inductor; Step S2: preprocessing the analog voltage signal to obtain a voltage waveform data segment; Step S3: performing multi-dimensional feature extraction on the voltage waveform data segment to obtain a voltage waveform multi-dimensional feature vector; Step S4: extracting voltage RMS effective value, total harmonic distortion rate and wavelet packet energy from the voltage waveform multi-dimensional feature vector; Step S5: inputting the voltage RMS effective value, total harmonic distortion rate and wavelet packet energy into a hierarchical event pre-detection model to obtain an event feature subset encoding vector; Step S6: passing the event feature subset encoding vector through an intelligent event fine classification model to obtain an accurate classification result of the power event.

[0028] It is worth mentioning that in step S1, the analog voltage signal collected by the mutual inductor is acquired. It can be understood that the analog voltage signal is just the physical carrier of the continuous change of the grid voltage in the time dimension, which completely contains all the waveform information of the grid in normal operation or when a disturbance occurs, whether it is a steady-state fundamental component or a dynamic harmonic, sag, swell, transient oscillation, etc., all in its most original and complete form. Therefore, acquiring the analog voltage signal is an absolute necessary prerequisite for any meaningful power quality analysis.

[0029] In particular, the analog voltage signal is collected by a transformer, which should be understood because in the actual power system, especially in the medium and high voltage grid, the voltage level is far from what electronic measuring devices can directly withstand. For example, the voltage of the distribution network can reach 10kV, and the voltage of the transmission network is even higher. It is extremely dangerous and unrealistic to directly connect the monitoring device to the high-voltage line. The voltage transformer plays a crucial role as a bridge. Its core function is to reduce the voltage on the high-voltage side by a precise transformation ratio, proportionally to a safe, low-voltage level that can be processed by measuring instruments, such as standard 100V or lower. More importantly, a high-quality voltage transformer can maintain the original voltage waveform shape and phase relationship as much as possible without distortion during voltage reduction. This fidelity is crucial for subsequent waveform analysis, and any distortion or phase shift introduced during voltage reduction will directly affect the accuracy of subsequent feature extraction, leading to errors in the final classification results. Therefore, collecting through the transformer is not only a safety consideration, but also a key technical means to ensure the accuracy of the source of measurement data.

[0030] In one embodiment, the analog voltage signal collected by the transformer includes: first, according to the voltage level of the monitoring point, select a voltage transformer with appropriate transformation ratio and accuracy level, and connect it in parallel to the high-voltage side of the power line to be monitored, such as phase A, phase B or phase C in a three-phase system. Then, connect the output terminals of the secondary side (low-voltage side) of the transformer to the analog input channel of the data acquisition module of the monitoring system through a dedicated signal cable with good shielding performance. This data acquisition module is a data acquisition card with a built-in high-speed, high-precision analog-to-digital converter or an integrated monitoring terminal. When the monitoring system starts, the software program will issue instructions to the data acquisition module through the drive interface, making it start to receive continuous analog voltage signals from the secondary side of the transformer.

[0031] Specifically, in step S2, the analog voltage signal is preprocessed to obtain a voltage waveform data segment. It should be understood that the original, continuous, and possibly noisy analog voltage signal is converted into a discrete, finite length, and clean digital sequence, i.e. voltage waveform data segment, so that it can be effectively processed, analyzed and feature extracted by subsequent computer programs. Without this preprocessing step, subsequent multi-dimensional feature extraction and intelligent classification model cannot be executed because digital computing systems cannot directly operate continuous analog quantities.

[0032] In one embodiment, the analog voltage signal is sampled at a predetermined sampling frequency to obtain a plurality of digital voltage sampling points; the plurality of digital voltage sampling points are low-pass filtered to obtain the voltage waveform data segment.

[0033] Specifically, the first step is sampling, and the second step is low-pass filtering. Sampling is the process of converting a continuous-time analog signal into a discrete-time digital sequence. In essence, it is to measure the instantaneous amplitude of the analog voltage signal at each predetermined, extremely small time interval point and convert it into a digital value. The key to performing this operation is the setting of the predetermined sampling frequency, which determines the density of sample points collected per unit time. According to the Nyquist-Shannon sampling theorem, in order to reconstruct the original signal without distortion, the sampling frequency must be at least twice the highest frequency component in the original signal. In power quality analysis, in order to capture high-frequency information such as high-order harmonics and transient disturbances, a sampling frequency much higher than twice the power frequency is usually used. After sampling, a raw data sequence composed of multiple digital voltage sample points is obtained. However, this raw sequence often contains high-frequency noise introduced by the measurement environment, the device itself, etc., which will interfere with the accuracy of subsequent feature analysis. Therefore, the second sub-step, low-pass filtering, is needed. Low-pass filtering is a signal processing technique that allows signal components below a certain cutoff frequency to pass through, while attenuating or blocking signal components above the cutoff frequency. In this application, a digital low-pass filter is applied to the sequence of digital voltage sample points obtained by sampling, and the purpose is to filter out high-frequency noise interference, while also effectively preventing frequency aliasing effects that may be caused by sampling, thereby obtaining a pure voltage waveform data segment that better reflects the true power quality condition.

[0034] In a specific embodiment, the pre-processing of the analog voltage signal to obtain a voltage waveform data segment includes: first, in the sampling link, a high-speed data acquisition card is used to sample the analog voltage signal obtained in the previous step at a predetermined sampling frequency of 12.8 kHz. For a 50 Hz power grid, a power frequency cycle is 20 milliseconds, and within this cycle, 256 digital voltage sample points can be collected at a frequency of 12.8 kHz. In order to perform effective time-domain and frequency-domain feature analysis, this embodiment selects to collect voltage data for 10 cycles continuously, so that an original data sequence containing 2560 sample points is obtained. Subsequently, in the low-pass filtering link, in order to eliminate the interference of high-frequency noise and aliasing effects, the original data sequence containing 2560 sample points is subjected to digital low-pass filtering processing. Specifically, a fourth-order Butterworth low-pass filter is used, with a cutoff frequency set to 2.5 kHz. The Butterworth filter is selected because of its frequency response characteristics of maximum flatness in the passband, which can ensure that the waveform in the effective signal frequency range does not produce additional distortion. The data sequence obtained after filtering is the final pure and formatted voltage waveform data segment used for subsequent multi-dimensional feature extraction.

[0035] More specifically, in step S3, multidimensional feature extraction is performed on the voltage waveform data segment to obtain a voltage waveform multidimensional feature vector. It should be understood that although the original voltage waveform data segment contains all the information about power quality, its data volume is large and the information is redundant. Directly inputting it into the subsequent classification model is not only computationally inefficient, but also difficult to effectively distinguish different types of power quality events. Therefore, through multidimensional feature extraction, the most relevant and discriminative information in the original waveform data related to the power quality status is condensed into a feature set with relatively low dimensionality and higher information density, namely the voltage waveform multidimensional feature vector. This feature vector can quantitatively describe the characteristics of the voltage waveform from different angles and levels, providing high-quality input for subsequent event pre-detection and fine classification. The multidimensional feature extraction refers to the calculation or extraction of multiple different types of parameters or indicators from the voltage waveform data segment, each of which characterizes a certain attribute of the waveform from a specific dimension. The voltage waveform multidimensional feature vector is a mathematical vector formed by arranging multiple extracted feature values ​​in a certain order. It constitutes a comprehensive and digital description of the power quality status of the current voltage waveform data segment.

[0036] like Figure 3 As shown, step S3 is specifically implemented as follows: Step S31: Calculating the voltage RMS value of the voltage waveform data segment; The RMS value of voltage is a key indicator for measuring the average level of voltage amplitude. Its calculation formula is: ; in, is the voltage value of the i-th sampling point in the voltage waveform data segment, is the total number of sampling points in the data segment, Indicates the RMS value of voltage. This RMS value can directly reflect amplitude disturbances such as voltage sag and swell.

[0037] Step S32: performing FFT analysis on the voltage waveform data segment to obtain the fundamental wave amplitude, the third harmonic amplitude, the fifth harmonic amplitude and the total harmonic distortion rate; Fast Fourier Transform (FFT) is an efficient algorithm that can convert voltage waveform data segments in the time domain into the frequency domain, thereby clearly revealing the amplitude and phase of each frequency component it contains. Through FFT analysis, the amplitude of the fundamental wave (usually 50Hz or 60Hz) and the amplitudes of typical low-order harmonics such as the third harmonic and the fifth harmonic can be directly read from the spectrum. Total harmonic distortion (THD) is a comprehensive indicator to measure the degree of waveform distortion, usually defined as the ratio of the root mean square value of all harmonic components to the root mean square value of the fundamental wave, that is, ; where, is the effective value of h-th harmonic, is the effective value of fundamental, H is the highest harmonic order considered, denotes the total harmonic distortion.

[0038] Step S33: extracting instantaneous features from the voltage waveform data segment to obtain wavelet packet energy; Unlike FFT which focuses on steady-state harmonic analysis, wavelet packet analysis can provide time-frequency localized information of the signal, and is very effective for capturing and characterizing transient disturbances in the voltage waveform such as voltage sag, swell, interruption, transient oscillation, spike, etc. By performing multi-level wavelet packet decomposition on the voltage waveform data segment, wavelet packet coefficients in different frequency bands can be obtained, and the energy (usually the sum of square of the coefficients) of the wavelet packet decomposition coefficients in a specific high-frequency band can be calculated to obtain wavelet packet energy features sensitive to transient events.

[0039] Step S34: extracting voltage flicker degree and voltage unbalance degree from the voltage waveform data segment; Voltage flicker degree is an index to evaluate the flickering effect of light caused by voltage fluctuation, and its calculation usually follows the algorithm specified in international standards (such as IEC 61000-4-15), involving statistical analysis of the voltage fluctuation envelope. Voltage unbalance degree is used to measure the degree of asymmetry in amplitude or phase of each phase voltage in a three-phase power system, and its calculation usually needs to combine the voltage data of the other two phases, for example, by calculating the ratio of negative sequence voltage component to positive sequence voltage component, or the ratio of the maximum deviation of phase voltage from average phase voltage to average phase voltage.

[0040] Step S35: arranging the voltage RMS effective value, fundamental amplitude, third harmonic amplitude, fifth harmonic amplitude, total harmonic distortion, wavelet packet energy, voltage flicker degree, and voltage unbalance degree into the voltage waveform multi-dimensional feature vector.

[0041] In one specific embodiment, the process of multi-dimension feature extraction on a voltage waveform data segment containing 2560 data points to obtain a voltage waveform multi-dimension feature vector includes: first, calculating the voltage RMS effective value: according to the 2560 sampling points in the voltage waveform data segment, the root mean square (RMS) effective value of the voltage within 10 cycles is calculated. Second, perform FFT analysis: apply fast Fourier transform (FFT) to the voltage waveform data segment to obtain its frequency spectrum information. The fundamental (50Hz) amplitude, third harmonic (150Hz) amplitude, and fifth harmonic (250Hz) amplitude are extracted from the frequency spectrum, and the total harmonic distortion (THD) is calculated according to the amplitude of the harmonic components. Third, extract instantaneous features: perform three-layer wavelet packet decomposition on the voltage waveform data segment, and select Daubechies 4 (db4) wavelet basis. By calculating the energy of the wavelet packet decomposition coefficients in a specific high-frequency band, the wavelet packet energy is obtained, which is very sensitive to voltage sag, swell, interruption and transient oscillation and other instantaneous disturbances. Fourth, extract other features: according to relevant international standards (such as IEC 61000-4-15), calculate the voltage flicker degree of the data segment; at the same time, combined with the voltage data of the other two phases (B phase and C phase), calculate the voltage unbalance degree. Finally, the 8 feature values calculated and extracted above, i.e. voltage RMS effective value, fundamental amplitude, third harmonic amplitude, fifth harmonic amplitude, total harmonic distortion, wavelet packet energy, voltage flicker degree and voltage unbalance degree, are sorted and combined into an 8-dimensional voltage waveform multi-dimension feature vector.

[0042] In step S4, the voltage RMS effective value, total harmonic distortion and wavelet packet energy are extracted from the voltage waveform multi-dimension feature vector. It should be understood that directly inputting the voltage waveform multi-dimension feature vector into an intelligent event fine classification model for analysis, although theoretically feasible, may bring unnecessary computational burden in real-time monitoring scenarios, and may affect the speed of preliminary judgment due to feature redundancy. Therefore, by selectively extracting the most representative and indicative core features, a lightweight and efficient pre-detection model can be built to quickly perform preliminary screening on power quality status and determine whether there is a significant disturbance event, laying a foundation for subsequent more detailed classification analysis.

[0043] Specifically, in the previous step (i.e., step S3), a multi-dimensional feature vector containing 8 characteristic values, including voltage RMS effective value, fundamental amplitude, third harmonic amplitude, fifth harmonic amplitude, total harmonic distortion, wavelet packet energy, voltage flicker degree, and voltage unbalance degree, has been completely constructed and stored. The operation of extracting the voltage RMS effective value, total harmonic distortion, and wavelet packet energy from the voltage waveform multi-dimensional feature vector is to accurately locate and extract three specific characteristic values from the 8-dimensional vector: the voltage RMS effective value, the total harmonic distortion, and the wavelet packet energy. In a specific embodiment, the 8-dimensional feature vector is stored as an array or list, and extracting the voltage RMS effective value, total harmonic distortion, and wavelet packet energy corresponds to accessing specific index positions of the array and copying the values at these three positions to form a new, lower-dimensional feature subset. The three selected features are given special status because they correspond to the three most common and core categories in power quality problems: the voltage RMS effective value mainly represents the stability of the voltage amplitude and is the primary indicator for judging voltage sag, voltage swell, and other amplitude disturbances; the total harmonic distortion (THD) is the gold standard for measuring the degree of deviation of the voltage waveform from the ideal sine wave, directly pointing to the harmonic pollution problem; and the wavelet packet energy, especially the high-frequency part of the wavelet packet energy, is extremely sensitive to transient disturbances with extremely short duration and concentrated energy (such as spikes and oscillations).

[0044] Further, in step S5, the voltage RMS effective value, total harmonic distortion, and wavelet packet energy are input into the hierarchical event pre-detection model to obtain an event feature subset encoding vector. It should be understood that through a lightweight preliminary diagnosis mechanism, it can quickly determine whether there is an abnormality in the current power quality and preliminarily classify the type of the abnormality. This hierarchical processing idea avoids using complex and detailed classification models for each data point, thereby greatly improving the response speed and computational efficiency of the real-time monitoring system and achieving a reasonable allocation of computing resources: only when a suspected event is preliminarily detected, does the subsequent more resource-intensive depth analysis start.

[0045] Specifically, the hierarchical event pre-detection model of the present application is not a complex machine learning model, but a high-efficiency rule judgment engine based on expert knowledge and engineering practice. It receives the three core features extracted in the previous step as input and makes judgments according to a series of pre-set logical rules. The event feature subset encoding vector output by the model is a dynamically generated vector for representing the core features of the suspected event. It is not a fixed-length vector, and its content and dimension completely depend on which rules in the pre-detection model are triggered. It is not only a preliminary encoding of the event type, but also a key subset extracted from the complete feature set.

[0046] AsFigure 4 As shown, step S5 is implemented as the following steps: Step S51: Determine whether the voltage RMS effective value is less than a preset threshold value. If the condition is met, trigger the suspected transient event rule to include the voltage RMS effective value in the event feature subset encoding vector. Step S52: Determine whether the total harmonic distortion rate is greater than a preset threshold value. If the condition is met, trigger the suspected harmonic event rule to include the total harmonic distortion rate in the event feature subset encoding vector. Step S53: Determine whether the wavelet packet energy is greater than a preset threshold value. If the condition is met, trigger the suspected transient event rule to include the wavelet packet energy in the event feature subset encoding vector.

[0047] In one specific embodiment, after inputting the voltage RMS effective value, total harmonic distortion rate, and wavelet packet energy into the hierarchical event pre-detection model, the model makes a judgment based on the following rules to generate the event feature subset encoding vector: First, the suspected transient event rule: determine whether the voltage RMS effective value is less than a preset threshold value V_th (for example, 198 V). If it is less than the threshold value, it is considered that a voltage sag event may have occurred, triggering this rule, and the voltage RMS effective value is included in the event feature subset encoding vector.

[0048] Second, the suspected harmonic event rule: determine whether the total harmonic distortion rate is greater than a preset threshold value THD_th (for example, 8%). If it is greater than the threshold value, it is considered that a harmonic over-limit event may have occurred, triggering this rule, and the total harmonic distortion rate is included in the event feature subset encoding vector.

[0049] Third, the suspected transient event rule: determine whether the wavelet packet energy is greater than an empirically set preset threshold value E_th (for example, 0.05). If it is greater than the threshold value, it is considered that a transient oscillation or spike event may have occurred, triggering this rule, and the wavelet packet energy is included in the event feature subset encoding vector. For example, if the voltage RMS effective value is monitored to be 185 V at a certain time, and the THD and wavelet packet energy are within the normal range, the final generated event feature subset encoding vector only contains [voltage RMS effective value]. If the THD is 10% and the others are normal, the vector is [total harmonic distortion rate]. If both the voltage RMS effective value and the THD are abnormal, the vector will be [voltage RMS effective value, total harmonic distortion rate]. This dynamically generated vector accurately encodes the initially diagnosed event information and is ready to be passed to the next step of the intelligent event fine classification model.

[0050] Further, in step S6, the event feature subset encoding vector is input into the intelligent event fine classification model to obtain the accurate classification result of the power quality event. It should be understood that the hierarchical event pre-detection model can efficiently identify waveform data segments suspected of having problems from a large amount of normal data, but it is essentially a coarse-grained filter, and its diagnostic ability is limited. First, it cannot effectively handle complex composite power quality events. For example, when a large nonlinear load is started, it may simultaneously cause voltage sag and serious harmonic distortion. At this time, the hierarchical event pre-detection model may trigger both the suspected transient event rule and the suspected harmonic event rule, generating an event feature subset encoding vector containing the voltage RMS effective value and the total harmonic distortion rate, but it cannot give an accurate composite event diagnosis conclusion such as voltage sag accompanied by harmonics. Secondly, the hierarchical event pre-detection model is also powerless to distinguish different sub-class events under the same large class. For example, transient oscillation and voltage spike may both cause significant increase in wavelet packet energy, but they differ greatly in waveform morphology, duration and impact on the system, and the hierarchical event pre-detection model cannot distinguish them by only using the wavelet packet energy as an indicator.

[0051] Finally, the hierarchical event pre-detection model completely ignores the rich information contained in other features in the multi-dimensional feature vector, such as fundamental amplitude, specific harmonic components, voltage flicker degree and unbalance degree, etc. These information is crucial for accurately understanding the cause and nature of the event. Therefore, in order to overcome these limitations and achieve comprehensive monitoring and accurate classification of power quality events, an intelligent event fine classification model that is more powerful, can comprehensively utilize all available information, and has deep analysis ability must be introduced. The model takes over the preliminary results of the hierarchical event pre-detection model and makes a second judgment.

[0052] As shown in Figure 5 , step S6 is implemented as the following steps: Step S61: eliminating the event feature subset encoding vector from the voltage waveform multi-dimensional feature vector to obtain a context feature vector; in the hierarchical event pre-detection stage, the core indicators triggering the alarm, i.e. the event feature subset encoding vector, have been identified, which can be regarded as the main symptoms of the event. In addition to these main symptoms, other features in the original 8-dimensional feature vector, such as fundamental amplitude, voltage flicker degree, etc., although not directly triggered by the threshold in the preliminary judgment, are not useless information, but constitute the context of the event occurrence. For example, a voltage sag (main symptom) occurs in the background of high three-phase voltage imbalance (context), and the causes and effects may be completely different from those of a sag occurring in the background of three-phase balance. By explicitly dividing the features into two sets of event features and context features, the application converts the classification problem from a simple multi-feature input, single-label output mode to a more advanced mode of analyzing how the context affects the event features and making judgments based on this. This separation enables the model to focus on learning and modeling the adjustment and influence of the context on the event, rather than mixing all features together for fuzzy correlation learning, thereby greatly improving the depth and accuracy of analysis.

[0053] Step S62: based on the context feature vector, event feature salientization is performed on the event feature subset encoding vector to obtain a salientized event feature subset encoding vector, so that the initially identified, relatively isolated event features become more significant, more informative and more easily distinguished by the final classifier after fully absorbing the context information of the environment in which they are located. This is not simply concatenating event features and context features, but generating a new, context information enhanced feature representation through a nonlinear interaction and reasoning process.

[0054] Step S63: inputting the salientized event feature subset encoding vector into a pre-trained Random Forest classification model to obtain the accurate classification result of the electric energy event.

[0055] Preferably, step S62 is implemented as the following steps: Step S621: Linearly transforming the event feature subset encoding vector to obtain a transformed event feature subset encoding vector; it should be understood that by a learnable linear transformation, the original event feature vector is projected from its inherent feature space to a new vector space that is more suitable for interaction analysis, that is, the dimension of the transformed event feature subset encoding vector is the same as that of the context feature vector. This means that a numerical value representing the depth of voltage sag, after this transformation, its new vector representation not only retains the original amplitude information, but is also mathematically adjusted to a state that can more effectively reveal its potential association with voltage flicker degree, harmonic content and other context features. This step does not increase or decrease information, but optimizes the form of information expression, which provides a normalized and semantically enhanced input for subsequent construction of a comprehensive and accurate feature interaction model.

[0056] Step S622: Calculating the event feature context global fine-grained interaction feature matrix of the transformed event feature subset encoding vector relative to the context feature vector, denoted as: ; wherein, denotes the transformed event feature subset encoding vector, denotes the context feature vector, denotes the transpose of the vector, denotes element-level multiplication, denotes position-wise addition, denotes matrix multiplication, denotes a dynamic weight matrix, denotes a projection matrix, the dynamic weight matrix and the dynamic weight matrix are obtained through training, denotes an activation function, denotes a hyperbolic tangent function, denotes an event feature context global fine-grained interaction feature matrix.

[0057] It should be appreciated that, by calculating the event feature context global fine-grained interaction feature matrix of the transformed event feature subset encoding vector relative to the context feature vector, the dimensionality of the analysis is raised from a one-dimensional feature list to a two-dimensional structured relationship graph. This step generates a matrix by calculating the interaction relationship of all element pairs between the two vectors, and each element of the matrix quantitatively describes the correlation strength between a certain specific aspect of the event feature and a certain specific aspect of the context feature. The technical effect is that it constructs a comprehensive and detailed numerical description of the interaction between features. For example, in a complex disturbance, the interaction matrix can simultaneously capture the strong positive correlation between the voltage sag depth and the third harmonic amplitude, and the weak correlation with the voltage unbalance degree. This global and fine-grained representation completely encapsulates the complex dependence between the event and its occurrence environment, forming a highly information-dense structured data entity. The two steps work together, and ultimately generate a global fine-grained interaction feature matrix that can fully reflect all potential interactions between event features and context features, which is the fundamental basis for subsequent sequential decomposition and time-dependent modeling.

[0058] Step S623: Matrix decompose the event feature context global fine-grained interaction feature matrix by column vector to obtain the sequence distribution of the event feature context local response encoding vector, denoted as: ; Wherein, represents the matrix decomposition operation, represents the sequence distribution of the event feature context local response encoding vector.

[0059] It should be appreciated that the global fine-grained interaction feature matrix encapsulates all the interactions between event features (e.g. voltage sag depth) and all context features (e.g. harmonic distortion, voltage unbalance, flicker, etc.). However, this concurrent representation fails to reveal the hierarchy or dependency of the influence of different context features on event features. By performing matrix decomposition on column vectors, this step transforms this complex global map into an ordered sequence. Each element in the sequence, i.e. an event feature-context local response encoding vector, precisely represents the collective response of all event feature dimensions to a certain specific context feature dimension. For example, the first vector in the sequence can represent the response of event features to harmonic distortion, the second vector represents the response to voltage unbalance, and so on. Meanwhile, by matrix decomposition, the model also introduces a structured analysis order, enabling the model to evaluate the cumulative effect of different context factors sequentially. When processing this sequence, the model can learn how the conditional dependency among context features jointly act on event features. For example, the model can learn that, given that the harmonic distortion is already high, the combined effect of voltage unbalance on event features is different from the effect of unbalance when the harmonic is normal. This sequential, cumulative modeling of interaction relationships enables the final generated feature representation to contain deeper, non-linear, and logically progressive semantic information, greatly enhancing the model's recognition accuracy and analytical ability for complex composite power quality events.

[0060] Step S624: inputting the sequence distribution of the event feature-context local response encoding vectors into an inference sequence derivation module based on a forward LSTM model to obtain the salient event feature subset encoding vector, denoted as: ; wherein, represents a long short-term memory network, represents a salient event feature subset encoding vector.

[0061] It should be appreciated that the inference sequence derivation module based on the forward long short-term memory network (LSTM) uses its unique gating mechanism and internal memory unit to perform ordered and cumulative information processing on the input local response encoding vector sequence. In the power quality monitoring scenario, after receiving the vector representing the event response to the harmonic distortion rate, the module updates its internal state; when it continues to receive subsequent vectors representing the response to the voltage unbalance degree, its processing mode will be adjusted by the previous harmonic response information. This sequential modeling of context feature dependency enables the model to accurately capture the unique combined effects of different disturbance factor combinations, such as distinguishing between a simple voltage sag and a complex sag event caused by the start of a large nonlinear load accompanied by harmonic distortion and voltage imbalance.

[0062] The final output of the salient event feature subset encoding vector is not a simple aggregation of the input sequence, but a highly abstracted semantic representation formed after learning deep temporal dependency and dynamically weighting context information. This vector compactly encodes the final integrated state after a series of context factors are gradually adjusted from the initial event features, and its internal structure contains a deep understanding of the entire event nature, thereby providing a highly discriminative decision basis for the subsequent classification model, significantly improving the recognition accuracy of complex power quality events.

[0063] Preferably, step S622 is specifically implemented as the following steps: Step S6221: Transpose multiplication of the transformed event feature subset encoding vector and the event feature subset encoding vector to obtain an event feature context association matrix; Step S6222: Determine a dynamic weight matrix and a projection matrix; perform element-level multiplication of the event feature context association matrix and the dynamic weight matrix, and pass it through a tanh function to obtain an event feature context asymmetric interaction representation matrix; Step S6223: Perform multiplication of the projection matrix and the matrix obtained by inputting the event feature context association matrix into a Sigmoid function for activation to obtain an event feature context gating matrix; Step S6224: Fuse the event feature context asymmetric interaction representation matrix and the event feature context gating matrix to obtain an event feature context global fine-grained interaction feature matrix.

[0064] Preferably, step S6224 is specifically implemented as: performing sequential projection consistency verification on the event feature context asymmetric interaction representation matrix and the event feature context gating matrix to obtain a transition matrix; let , and , the event feature context asymmetric interaction representation matrix and the event feature context gating matrix Perform sequential projection consistency verification to obtain the transition matrix ,satisfy: ; in, Represents the event feature context asymmetric interaction representation matrix, represents the event feature context gating matrix, represents the transition matrix, Represents matrix multiplication.

[0065] That is, it is considered that different projection parts have an interactive relationship with each other, so that the transition matrix Based on different interaction orders, there are corresponding generator representations and inverse representations.

[0066] The transition matrix is ​​subjected to a distribution state degeneration operation to obtain a correction matrix, namely: ; in, represents the transition matrix, represents the inverse matrix of the transition matrix, Represented by the transition matrix Each eigenvalue of is the power function matrix of the exponential, represents the F norm of the matrix, represents the correction matrix.

[0067] That is, the transition matrix and its inverse matrix As the group representation of the generator representation and the inverse element representation, respectively, through the matrix The norm is c, that is, the distribution state group representation is irreducibly degenerated, so as to obtain the interaction relationship representation based on the ground state. In this way, the unidirectionality of the generator representation and the inverse element representation can be realized based on the transition difference representation between the ground states.

[0068] Finally, based on the correction matrix To correct the event feature context asymmetric interaction representation matrix and the event feature context gating matrix ,Right now , Then, the event feature context global fine-grained interaction feature matrix is ​​obtained by adding them up according to the position, so as to improve the fusion expression effect of the event feature context global fine-grained interaction feature matrix.

[0069] Particularly, in the calculation of the event feature context global fine-grained interaction feature matrix, since the event feature context correlation matrix respectively adopts a dynamic weight matrix and a projection matrix to modulate in different directions, and the functions and have different orders of exponential functions based on natural constants, which, while ensuring comprehensive coverage of the global interaction, also leads the fine-grained subtle semantic representation to different interaction paths, thereby affecting the fusion effect of the two parts.

[0070] Therefore, in order to realize the equivalence of the interaction path, it is expected to ensure the invariance of the correlation feature space based on different interaction paths, that is, to improve the constraint robustness of the correlation feature space for different interaction paths, thereby improving the fusion effect of the fine-grained subtle semantic representation.

[0071] For step S63, finally, the significant event feature subset encoding vector is input into the pre-trained Random Forest classification model to obtain the accurate classification result of the power event. After the foregoing feature significant process, a significant event feature subset encoding vector is obtained, and at this time a powerful and robust classifier is needed for classification. The present application chooses to use the Random Forest model, and this choice is based on its recognized advantages in the field of machine learning. Random Forest is an ensemble learning algorithm that does not rely on a single model for decision-making, but builds a forest composed of numerous decision trees, and the final classification result is determined by the voting results of all decision trees in the forest. The reason for choosing it is that it has extremely strong anti-overfitting ability. Since each tree is trained on different data subsets (obtained by Bootstrap sampling), and only part of the features is considered when splitting at each node, the trees in the forest have diversity, thereby avoiding the defect that a single decision tree is prone to overfitting the training data. Secondly, random forest performs well on high-dimensional feature data and can effectively process the significant event feature subset encoding vector generated, which may have high dimensions and complex relationships between features. Thirdly, it is not sensitive to noise and outliers in the data, which is a great advantage when dealing with real power quality data full of interference. Finally, the random forest model has relatively fast training speed and usually achieves very good performance on various classification tasks.

[0072] Specifically, the architecture of the pre-trained Random Forest classification model is as follows: its basic unit is a decision tree. Each decision tree is a tree-shaped prediction model, where each internal node represents a test on a feature (here, a dimension in the feature vector), each branch represents a test output, and each leaf node represents a final power event category (such as "voltage sag"). In building this forest, the present application employs two core randomization techniques: first, data randomization, i.e., by Bootstrap self-sampling method, a sample is drawn with replacement from the original training dataset to generate a unique training subset for each tree in the forest. Second, feature randomization, i.e., when building each node of each tree, instead of searching for the optimal split point among all features, a randomly selected feature subset is selected. When a new feature vector of a subset of significant events needs to be classified, the vector is input into each decision tree in the forest. Each tree will independently give a classification result, which is like each expert in the forest giving his own opinion. Finally, the random forest model will adopt the majority voting principle to count all the classification results of the trees, and the class with the most votes will be output as the final accurate classification result of the power event. Here, this complex model containing LSTM and random forest has been fully offline trained and optimized on a large dataset containing a large number of power quality event samples with accurate labels before being deployed to the field for real-time monitoring, so that the internal parameters (such as linear transformation matrix, LSTM weight, random forest structure, etc.) are in the optimal state, thereby ensuring the accuracy and reliability of online classification.

[0073] As shown in Figure 6 The application further provides a power quality comprehensive monitoring system 100 based on real-time waveform analysis, comprising: an analog voltage signal acquisition module 110, configured to acquire an analog voltage signal collected by a mutual inductor; an analog voltage signal preprocessing module 120, configured to preprocess the analog voltage signal to obtain a voltage waveform data segment; a multi-dimensional feature extraction module 130, configured to perform multi-dimensional feature extraction on the voltage waveform data segment to obtain a voltage waveform multi-dimensional feature vector; a voltage waveform multi-dimensional feature analysis module 140, configured to extract a voltage RMS effective value, a total harmonic distortion rate and a wavelet packet energy from the voltage waveform multi-dimensional feature vector; an event feature pre-detection and coding module 150, configured to input the voltage RMS effective value, the total harmonic distortion rate and the wavelet packet energy into a hierarchical event pre-detection model to obtain an event feature subset coding vector; The event feature coding classification analysis module 160 is configured to pass the event feature subset coding vector through an intelligent event fine classification model to obtain an electric energy event accurate classification result.

[0074] It is worth mentioning that the technical features such as transformers involved in the present patent application should be regarded as prior art, and the specific structure, working principle and possible control mode and spatial arrangement mode of these technical features can be selected conventionally in the art, and should not be regarded as the invention point of the present patent, and the present patent will not be further specifically expanded and described.

[0075] For those skilled in the art, the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced equivalently, and any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A power quality comprehensive monitoring method based on real-time waveform analysis, characterized in that, The method comprises the following steps: Step S1: acquiring an analog voltage signal collected by a mutual inductor; Step S2: preprocessing the analog voltage signal to obtain a voltage waveform data segment; Step S3: performing multi-dimensional feature extraction on the voltage waveform data segment to obtain a voltage waveform multi-dimensional feature vector; Step S4: extracting a voltage RMS effective value, a total harmonic distortion rate and a wavelet packet energy from the voltage waveform multi-dimensional feature vector; Step S5: inputting the voltage RMS effective value, the total harmonic distortion rate and the wavelet packet energy into a hierarchical event pre-detection model to obtain an event feature subset encoding vector; Step S6: passing the event feature subset encoding vector through an intelligent event fine classification model to obtain an accurate classification result of the electric energy event.

2. The power quality comprehensive monitoring method based on real-time waveform analysis according to claim 1, characterized in that, In step S2, the analog voltage signal is sampled at a predetermined sampling frequency to obtain a plurality of digital voltage sampling points; the plurality of digital voltage sampling points are low-pass filtered to obtain the voltage waveform data segment.

3. The method for comprehensive monitoring of power quality based on real-time waveform analysis according to claim 1, characterized in that, Step S3 is specifically implemented as the following steps: Step S31: calculating a voltage RMS effective value of the voltage waveform data segment; Step S32: performing FFT analysis on the voltage waveform data segment to obtain a fundamental amplitude, a third harmonic amplitude, a fifth harmonic amplitude and a total harmonic distortion rate; Step S33: extracting an instantaneous feature from the voltage waveform data segment to obtain a wavelet packet energy; Step S34: extracting a voltage flicker degree and a voltage unbalance degree from the voltage waveform data segment; Step S35: arranging the voltage RMS effective value, the fundamental amplitude, the third harmonic amplitude, the fifth harmonic amplitude, the total harmonic distortion rate, the wavelet packet energy, the voltage flicker degree and the voltage unbalance degree into the voltage waveform multi-dimensional feature vector.

4. The power quality comprehensive monitoring method based on real-time waveform analysis according to claim 3, characterized in that, Step S5 is specifically implemented as the following steps: Step S51: judging whether the voltage RMS effective value is less than a preset threshold value, if the condition is met, triggering a suspected transient event rule to include the voltage RMS effective value in the event feature subset encoding vector; Step S52: judging whether the total harmonic distortion rate is greater than a preset threshold value, if the condition is met, triggering a suspected harmonic event rule to include the total harmonic distortion rate in the event feature subset encoding vector; Step S53: judging whether the wavelet packet energy is greater than a preset threshold value, if the condition is met, triggering a suspected transient event rule to include the wavelet packet energy in the event feature subset encoding vector.

5. The method for comprehensive monitoring of power quality based on real-time waveform analysis according to claim 4, characterized in that, Step S6 is specifically implemented as the following steps: Step S61: excluding the event feature subset encoding vector from the voltage waveform multi-dimensional feature vector to obtain a context feature vector; Step S62: based on the context feature vector, performing event feature saliency on the event feature subset encoding vector to obtain a salient event feature subset encoding vector; Step S63: inputting the salient event feature subset encoding vector into a pre-trained Random Forest classification model to obtain the accurate classification result of the electric energy event.

6. The method for comprehensive monitoring of power quality based on real-time waveform analysis according to claim 5, characterized in that, Step S62 is specifically implemented as the following steps: Step S621: performing linear transformation on the event feature subset encoding vector to obtain a transformed event feature subset encoding vector; Step S622: calculate the event feature context global fine-grained interaction feature matrix of the transformed event feature subset encoding vector relative to the context feature vector; Step S623: matrix decomposition is performed on the event feature context global fine-grained interaction feature matrix by column vector to obtain the sequence distribution of the event feature context local response encoding vector; Step S624: input the sequence distribution of the event feature context local response encoding vector into the inference sequence derivation module based on the forward LSTM model to obtain the salient event feature subset encoding vector.

7. The method for comprehensive monitoring of power quality based on real-time waveform analysis according to claim 6, characterized in that, For step S622, the following steps are implemented: Step S6221: multiply the transformed event feature subset encoding vector and the transposed event feature subset encoding vector to obtain an event feature context association matrix; Step S6222: determine a dynamic weight matrix and a projection matrix; perform element-level multiplication on the event feature context association matrix and the dynamic weight matrix, and pass it through a tanh function to obtain an event feature context asymmetric interaction representation matrix; Step S6223: multiply the projection matrix and the event feature context association matrix after activation by the Sigmoid function to obtain an event feature context gating matrix; Step S6224: fuse the event feature context asymmetric interaction representation matrix and the event feature context gating matrix to obtain an event feature context global fine-grained interaction feature matrix.

8. The power quality comprehensive monitoring method based on real-time waveform analysis according to claim 7, characterized in that, Step S6224 is implemented as: sequentially project the event feature context asymmetric interaction representation matrix and the event feature context gating matrix to obtain a transition matrix; perform distribution state degeneration operation on the transition matrix to obtain a correction matrix; Based on the correction matrix, the event feature context asymmetric interaction representation matrix and the event feature context gating matrix are respectively corrected, and then added by position to obtain an event feature context global fine-grained interaction feature matrix.

9. A power quality comprehensive monitoring system based on real-time waveform analysis, applied to the power quality comprehensive monitoring method based on real-time waveform analysis of any one of claims 1-8, characterized in that, Comprise: An analog voltage signal acquisition module configured to acquire an analog voltage signal collected by a mutual inductor; An analog voltage signal preprocessing module configured to preprocess the analog voltage signal to obtain a voltage waveform data segment; A multi-dimensional feature extraction module configured to perform multi-dimensional feature extraction on the voltage waveform data segment to obtain a voltage waveform multi-dimensional feature vector; A voltage waveform multi-dimensional feature analysis module configured to extract a voltage RMS effective value, a total harmonic distortion rate, and a wavelet packet energy from the voltage waveform multi-dimensional feature vector; An event feature pre-detection and encoding module configured to input the voltage RMS effective value, the total harmonic distortion rate, and the wavelet packet energy into a hierarchical event pre-detection model to obtain an event feature subset encoding vector; An event feature encoding classification analysis module configured to pass the event feature subset encoding vector through an intelligent event fine classification model to obtain an electric energy event accurate classification result.

Images