A Power Quality Monitoring Method and Device Based on AI Engine Optimization

Abstract

This invention discloses a power quality monitoring method and apparatus based on an AI engine optimization. The method includes the following steps: S1: acquiring voltage waveform data collected by monitoring equipment to obtain the sequence distribution of short-time window data of the voltage waveform; S2: obtaining the sequence distribution of structured encoding vectors of key power quality parameters; S3: obtaining the time-series context encoding vectors of key power quality parameters; S4: generating reconstructed voltage waveform data; and S5: determining whether there are any abnormal precursors based on the reconstruction error. The power quality monitoring method and apparatus based on an AI engine optimization disclosed in this invention utilizes the learning ability of AI models on waveform characteristics of normal power conditions to achieve early warning of power quality disturbance events, facilitating preventative maintenance and proactive intervention.

Description

Technical Field

[0001] This invention belongs to the field of power quality monitoring technology, specifically relating to a power quality monitoring method and a power quality monitoring device based on AI engine optimization. Background Technology

[0002] Power quality is a crucial indicator for evaluating the quality of power supply in a power system, directly impacting the safe and stable operation of the power system and the efficiency and lifespan of user equipment. With the increasing demands for power supply reliability and purity from modern industry, commerce, and information technology, power quality issues such as voltage sags, voltage rises, harmonics, frequency deviations, and three-phase imbalances can lead to malfunctions in sensitive equipment, production interruptions, equipment damage, and even safety accidents, causing significant economic losses. Therefore, developing efficient and accurate power quality monitoring solutions to promptly detect and warn of potential power quality disturbances is of paramount importance for ensuring grid security, improving power supply service levels, and protecting user interests.

[0003] Traditional power quality monitoring solutions typically rely on dedicated monitoring equipment deployed at critical nodes to measure and record various power quality parameters according to preset standards (such as IEC 61000-4-30). However, these methods often focus on recording and classifying disturbances that have occurred and reached defined thresholds, resulting in significant storage and transmission challenges and heavy computational burdens when dealing with massive amounts of continuous voltage waveform data. More importantly, traditional methods have limited ability to capture subtle, gradual anomalies preceding disturbances, making it difficult to effectively utilize the complex temporal correlation information contained in the data for early warning. Detection often occurs only after the problem has significantly worsened, limiting the possibility of preventative maintenance and proactive intervention.

[0004] Therefore, further improvements will be made to address the aforementioned issues. Summary of the Invention

[0005] The main objective of this invention is to provide a power quality monitoring method and apparatus based on an AI engine optimization. This method segments voltage waveform data into short-time windows and extracts multiple key power quality parameters from the voltage waveform data of each window, constructing a time-series distribution of multidimensional power quality parameters to characterize the evolution of the power state. Next, a deep learning algorithm is introduced to extract features and model the temporal context of the key power quality parameters in each short-time window, learning the temporal evolution law of the power quality state. Based on this, the prior knowledge learned by the decoder under normal power quality data is used to reconstruct the voltage waveform data. Then, based on the reconstruction error between the reconstructed and original voltage waveform data, abnormal precursors are identified. This method utilizes the learning ability of the AI ​​model on the waveform characteristics of normal power state to achieve early warning of power quality disturbance events, facilitating preventative maintenance and proactive intervention.

[0006] To achieve the above objectives, this invention provides a power quality monitoring method based on AI engine optimization, comprising the following methods:

[0007] Step S1: Acquire voltage waveform data collected by the (PQ) monitoring device, and divide the voltage waveform data into short-time windows to obtain the sequence distribution of the short-time window data of the voltage waveform;

[0008] Step S2: Extract the power quality parameter sequence of each voltage waveform short-time window data from the sequence distribution of the voltage waveform short-time window data to obtain the sequence distribution of the structured encoding vector of the key power quality parameters;

[0009] Step S3: Perform context coding based on temporal adversarial repulsion guidance on the sequence distribution of the structured coding vector of the key power quality parameters to obtain the temporal context coding vector of the key power quality parameters;

[0010] Step S4: Generate reconstructed voltage waveform data based on the timing context encoding vector of the key power quality parameters;

[0011] Step S5: Based on the reconstruction error between the voltage waveform data and the reconstructed voltage waveform data, determine whether there are any abnormal precursors.

[0012] As a further preferred technical solution of the above technical solution, in step S2, the power quality parameter sequence of each voltage waveform short-time window data in the sequence distribution is extracted to obtain the sequence distribution of the key power quality parameter sequence; the key power quality parameter embedding coding matrix is ​​used to perform structured coding on each key power quality parameter sequence in the sequence distribution of the key power quality parameter sequence to obtain the sequence distribution of the key power quality parameter structured coding vector.

[0013] As a further preferred technical solution to the above technical solution, step S3 is specifically implemented as follows:

[0014] Step S3.1: Extract the current key power quality parameter structured encoding vector from the sequence distribution of the key power quality parameter structured encoding vectors, and define other key power quality parameter structured encoding vectors in the sequence distribution of the key power quality parameter structured encoding vectors as key power quality parameter structured encoding vectors to be propagated, so as to obtain the sequence distribution of key power quality parameter structured encoding vectors to be propagated;

[0015] Step S3.2: Based on the temporal correlation between each structured encoding vector of the key power quality parameters to be propagated in the sequence distribution of the structured encoding vector of the key power quality parameters to be propagated and the current structured encoding vector of the key power quality parameters, feature propagation modulation is performed on each structured encoding vector of the key power quality parameters to be propagated in the sequence distribution of the structured encoding vector of the key power quality parameters to be propagated, so as to obtain a set of dynamic encoding vectors of key power quality parameter messages;

[0016] Step S3.3: Merge the set of dynamic encoding vectors of the key power quality parameter messages and the current structured encoding vector of the key power quality parameter to obtain the timing context encoding vector of the key power quality parameter.

[0017] As a further preferred technical solution to the above technical solution, step S3.2 is specifically implemented as follows:

[0018] Step S3.2.1: Calculate the dynamic transmission potential energy of each structured encoding vector of the key power quality parameter to be propagated relative to the current structured encoding vector in the sequence distribution of the structured encoding vector of the key power quality parameter to be propagated, so as to obtain the sequence distribution of the dynamic transmission potential energy encoding vector of the key power quality parameter node to be propagated;

[0019] Step S3.2.2: Calculate the message propagation repulsion coefficient of each structured encoding vector of the key power quality parameter to be propagated relative to the current structured encoding vector of the key power quality parameter in the sequence distribution of the structured encoding vector of the key power quality parameter to be propagated, so as to obtain the sequence distribution of the message propagation repulsion coefficient of the key power quality parameter node;

[0020] Step S3.2.3: Based on the sequence distribution of the repulsion coefficient of the node message propagation of the key power quality parameter to be propagated and the sequence distribution of the dynamic transmission potential energy encoding vector of the node message propagation of the key power quality parameter to be propagated, perform dynamic adaptive message propagation encoding on each structured encoding vector of the key power quality parameter to be propagated in the sequence distribution of the structured encoding vector of the key power quality parameter to be propagated, so as to obtain a set of dynamic encoding vectors of key power quality parameter messages;

[0021] The specific implementation of step S3.3 is as follows:

[0022] Step S3.3.1: Calculate the positional summation of the set of dynamic coding vectors for the key power quality parameter messages to obtain the dynamic propagation aggregate coding vector for the key power quality parameter messages;

[0023] Step S3.3.2: Perform weighted fusion of the dynamic propagation aggregation encoding vector of the key power quality parameter message and the current key power quality parameter structured encoding vector to obtain the key power quality parameter temporal context encoding vector.

[0024] As a further preferred technical solution to the above technical solution, step S3.3.2 is specifically implemented as follows:

[0025] Dynamic-static feature interaction sensing is performed on the dynamic propagation aggregation coding vector of the key power quality parameter message and the current key power quality parameter structured coding vector to obtain a fusion weight coefficient. The fusion weight coefficient is used as the weight of the current key power quality parameter structured coding vector, and the difference between the fusion weight coefficient and the current key power quality parameter structured coding vector is used as the weight of the dynamic propagation aggregation coding vector of the key power quality parameter message. The dynamic propagation aggregation coding vector of the key power quality parameter message and the current key power quality parameter structured coding vector are weighted and fused to obtain the key number time-series context coding vector.

[0026] As a further preferred technical solution of the above technical solution, in step S4, the key power quality parameter timing context encoding vector is subjected to feature renormalization based on power quality parameter node collaboration to obtain an optimized key power quality parameter timing context encoding vector; the optimized key power quality parameter timing context encoding vector is input into the voltage waveform reconstruction module based on the decoder to obtain the reconstructed voltage waveform data.

[0027] As a further preferred technical solution of the above technical solution, in step S5, the reconstruction error is the root mean square error between the voltage waveform data and the reconstructed voltage waveform data; based on the comparison between the reconstruction error and a preset threshold, it is determined whether there are any abnormal precursors.

[0028] To achieve the above objectives, the present invention also provides a power quality monitoring device based on AI engine optimization, comprising:

[0029] The voltage waveform data acquisition module is used to acquire voltage waveform data collected by the monitoring equipment.

[0030] A voltage waveform short-time window segmentation module is used to segment the voltage waveform data into short-time windows to obtain a sequence distribution of the voltage waveform short-time window data;

[0031] The power quality feature extraction module is used to extract the power quality features of each voltage waveform short-time window data in the sequence distribution of the voltage waveform short-time window data, so as to obtain the sequence distribution of the structured encoding vector of key power quality parameters;

[0032] The power quality parameter context encoding module is used to perform context encoding based on temporal adversarial repulsion guidance on the sequence distribution of the structured encoding vector of the key power quality parameters to obtain the temporal context encoding vector of the key power quality parameters.

[0033] The reconstructed voltage waveform data generation module is used to generate reconstructed voltage waveform data based on the timing context encoding vector of the key power quality parameters.

[0034] An abnormal precursor monitoring module is used to determine whether there are abnormal precursors based on the reconstruction error between the voltage waveform data and the reconstructed voltage waveform data. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the device of the present invention. Detailed Implementation

[0036] The following description is intended to disclose the present invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art. The basic principles of the invention defined in the following description can be applied to other embodiments, modifications, improvements, equivalents, and other technical solutions that do not depart from the spirit and scope of the invention.

[0037] In the preferred embodiments of the present invention, those skilled in the art should note that the PQ monitoring equipment and the like involved in the present invention can be considered as prior art.

[0038] Preferred embodiment.

[0039] This invention discloses a power quality monitoring method based on AI engine optimization, comprising the following methods:

[0040] Step S1: Acquire voltage waveform data collected by the (PQ) monitoring device, and divide the voltage waveform data into short-time windows to obtain the sequence distribution of the short-time window data of the voltage waveform;

[0041] Step S2: Extract the power quality parameter sequence of each voltage waveform short-time window data from the sequence distribution of the voltage waveform short-time window data to obtain the sequence distribution of the structured encoding vector of the key power quality parameters;

[0042] Step S3: Perform context coding based on temporal adversarial repulsion guidance on the sequence distribution of the structured coding vector of the key power quality parameters to obtain the temporal context coding vector of the key power quality parameters;

[0043] Step S4: Generate reconstructed voltage waveform data based on the timing context encoding vector of the key power quality parameters;

[0044] Step S5: Based on the reconstruction error between the voltage waveform data and the reconstructed voltage waveform data, determine whether there are any abnormal precursors.

[0045] For step S1, firstly, voltage waveform data collected by the PQ monitoring equipment is acquired. It should be understood that traditional power quality monitoring relies on dedicated equipment at fixed nodes, limiting data acquisition density and real-time performance, and the massive amount of continuous waveform data leads to excessively high storage and transmission costs. Therefore, to capture high-frequency disturbance characteristics in the power grid in real time and reduce data redundancy, this invention is based on the distributed deployment of high-precision PQ monitoring equipment. It synchronously acquires the raw three-phase voltage waveform signals at a high sampling rate (e.g., above 10 kHz), covering the dynamic changes of all nodes in the power grid. Specifically, intelligent PQ monitoring equipment supporting IEC standards is installed at key nodes in the power grid (such as substations and important load access points). Its high-speed ADC module synchronously samples and digitizes the voltage signal, generating a voltage waveform sequence containing timestamps, which is then locally cached and preprocessed through edge computing nodes. In this way, the detailed characteristics of the voltage waveform (such as high-frequency harmonics and transient drops) are preserved, while avoiding the bandwidth pressure of traditional centralized storage, providing a high-quality data foundation for subsequent analysis.

[0046] Secondly, considering the characteristics of continuous voltage waveform data—long duration and large data volume—direct global analysis would lead to an exponential increase in computational complexity and make it difficult to capture the precursor features of short-term disturbances. Therefore, to reduce computational load while preserving the local details of the voltage waveform's temporal sequence, this invention uses a sliding window mechanism to segment the continuous voltage waveform data into multiple short-time overlapping windows (e.g., each window is 200ms long with a 30% overlap rate), thus obtaining a sequence distribution of the short-time window data. It should be understood that precursors to power quality anomalies often manifest as short-duration (millisecond-level) local distortions with weak amplitude changes. Short-time windows can isolate such events and avoid the loss of key information through window overlap. In specific implementation, a Hamming window function can be used to window the original waveform to suppress spectral leakage, while phase alignment between windows ensures temporal continuity. In this way, the original voltage waveform data is converted into a window sequence with a time index, reducing the data size of a single processing operation and providing structured input for subsequent temporal modeling.

[0047] Specifically, in step S2, the power quality parameter sequences of each voltage waveform short-time window data in the sequence distribution are extracted to obtain the sequence distribution of the key power quality parameter sequences; the key power quality parameter embedding coding matrix is ​​used to perform structured coding on each key power quality parameter sequence in the sequence distribution of the key power quality parameter sequences to obtain the sequence distribution of the key power quality parameter structured coding vector.

[0048] Because the original voltage waveform contains multi-dimensional physical characteristics (such as amplitude, frequency, and harmonics), directly inputting it into the model would lead to feature redundancy and difficulty in interpretation. Therefore, in order to construct a quantifiable and interpretable power state characterization system, this invention, based on IEC standards and domain knowledge, extracts multiple key power quality parameters from short-time window data of each voltage waveform through parallel computing, constructing a key power quality parameter sequence for each short-time window to comprehensively reflect the spatiotemporal evolution characteristics of the voltage waveform. Specifically, the key power quality parameter sequence includes the effective voltage value, voltage deviation, crest factor, total harmonic distortion rate, harmonic content, frequency deviation, and voltage imbalance, wherein:

[0049] The effective voltage value (RMS) reflects the average level of voltage fluctuations;

[0050] Voltage deviation (percentage of nominal value) describes the degree of steady-state deviation between the actual voltage and the rated voltage;

[0051] The crest factor (peak value / RMS) reveals the ratio of the peak value to the effective value of the voltage waveform, which is of great significance for assessing instantaneous waveform distortion.

[0052] Total harmonic distortion (THD) measures the impact of harmonic components on the fundamental frequency and is a key indicator for assessing the purity of electrical energy.

[0053] The content of each harmonic (e.g., 3rd, 5th, 7th) provides detailed information about the harmonic spectrum, which helps to identify specific harmonic sources;

[0054] Frequency deviation (the difference from the 50Hz / 60Hz reference value) reflects the difference between the actual power grid frequency and the nominal frequency, and is crucial for power grid stability and equipment compatibility.

[0055] Voltage imbalance (proportion of negative sequence components) is used to quantify the degree of imbalance in a three-phase voltage system, which is particularly crucial for protecting the normal operation of three-phase equipment.

[0056] By performing parallel extraction of the aforementioned key power quality parameters from short-time window data of various voltage waveforms, high-dimensional waveform data is compressed into low-dimensional feature vectors. The sequence distribution of the key power quality parameter sequence is constructed as a "feature fingerprint" of power quality evolution. This not only preserves anomaly indicators with clear physical meaning and enhances the interpretability of anomaly detection, but also simplifies model input and provides an accurate feature basis for subsequent anomaly detection.

[0057] Considering the heterogeneity (differences in dimensions and numerical ranges) of the multivariate parameters in the key power quality parameter sequence, directly using the concatenated sequence of multivariate parameters as input to an AI model would make it difficult for the model to learn the implicit relationships between parameters. Therefore, in order to map multivariate physical parameters to a unified feature space and capture cross-parameter coupling relationships, this invention, based on the principle of representation learning, first standardizes each parameter in the key power quality parameter sequence to eliminate the influence of dimensions. Then, a trainable key power quality parameter embedding encoding matrix is ​​constructed using a deep learning framework. The standardized key power quality parameter sequences are then structured and encoded. The key power quality parameter vector of each short-time window is mapped to a high-dimensional embedding vector through matrix multiplication, resulting in the sequence distribution of the key power quality parameter structured encoding vectors. Simultaneously, backpropagation optimization is used to make the distance between the key power quality parameter structured encoding vectors in the embedding space reflect the correlation of power quality states in each short-time window, thereby capturing the dynamic evolution law of power quality states.

[0058] More specifically, since voltage waveform data is essentially time-series data, it contains rich information on the historical evolution of voltage waveforms and clues for predicting future trends. Therefore, in order to fully explore the time-series dependencies and abnormal evolution patterns in voltage waveform data and enhance the model's sensitivity to key abnormal events, this invention proposes a context encoding method based on time-series adversarial repulsion guidance. This method simulates the concepts of potential energy and repulsion in physics, treating the structured encoding vectors of key power quality parameters at historical time nodes as "sources" with different influences on the current state. Based on the distance (time interval) between historical nodes and the current node and their correlation with power quality state, the propagation potential energy and repulsion of the structured encoding vectors of key power quality parameters at each historical node are calculated, thereby dynamically modulating them to enhance the model's sensitivity to abnormal patterns. Through the aggregation of contextual information from global nodes, a time-series context encoding vector of key power quality parameters that integrates historical evolution patterns and clues for predicting future trends is obtained, providing a more refined feature representation for subsequent voltage waveform data reconstruction. Step S3 is specifically implemented as follows:

[0059] Step S3.1: Extract the current key power quality parameter structured encoding vector from the sequence distribution of the key power quality parameter structured encoding vectors, and define other key power quality parameter structured encoding vectors in the sequence distribution of the key power quality parameter structured encoding vectors as key power quality parameter structured encoding vectors to be propagated, so as to obtain the sequence distribution of key power quality parameter structured encoding vectors to be propagated, specifically, expressed by the formula:

[0060] ;

[0061] ;

[0062] ;

[0063] in, It is the sequence distribution of the structured encoded vectors of the key power quality parameters. It is in the sequence distribution Structured encoding vectors of key power quality parameters Represents the set of real numbers. It is the number of structured coding vectors for key power quality parameters. This indicates the length of the structured encoding vector for key power quality parameters. This represents the structured encoded vector of current key power quality parameters. This represents the sequence distribution of the structured encoded vectors of the key power quality parameters to be propagated.

[0064] Step S3.2: Based on the temporal correlation between each structured encoding vector of the key power quality parameters to be propagated in the sequence distribution of the structured encoding vector of the key power quality parameters to be propagated and the current structured encoding vector of the key power quality parameters, feature propagation modulation is performed on each structured encoding vector of the key power quality parameters to be propagated in the sequence distribution of the structured encoding vector of the key power quality parameters to be propagated, so as to obtain a set of dynamic encoding vectors of key power quality parameter messages;

[0065] Step S3.3: Merge the set of dynamic encoding vectors of the key power quality parameter messages and the current structured encoding vector of the key power quality parameter to obtain the timing context encoding vector of the key power quality parameter.

[0066] Furthermore, step S3.2 is specifically implemented as follows:

[0067] Step S3.2.1: Calculate the dynamic transmission potential energy of each structured encoding vector of the key power quality parameter to be propagated relative to the current structured encoding vector in the sequence distribution of the structured encoding vector of the key power quality parameter to be propagated, so as to obtain the sequence distribution of the dynamic transmission potential energy encoding vector of the key power quality parameter node to be propagated. Specifically, it is expressed by the formula:

[0068] ;

[0069] ;

[0070] ;

[0071] in, This represents the first weight matrix during pre-training. This represents the pre-trained second weight matrix. This represents the pre-trained weight parameters. Represents matrix multiplication. , and These represent dot multiplication, dot addition, and dot subtraction, respectively. The sequence distribution of the structured encoded vectors representing the key power quality parameters to be propagated. Structured encoding vectors of key power quality parameters for Compared to The dynamic transmission potential energy encoding vector of the key power quality parameter node to be propagated. This represents calculating the square of the eigenvalues ​​at each position in the vector;

[0072] Step S3.2.1: Calculate the message propagation repulsion coefficient of each structured encoding vector of the key power quality parameter to be propagated relative to the current structured encoding vector of the key power quality parameter in the sequence distribution of the structured encoding vector of the key power quality parameter to be propagated, so as to obtain the sequence distribution of the message propagation repulsion coefficient of the key power quality parameter node, specifically, expressed by the formula:

[0073] ;

[0074] in, For timestamp extraction function, Represents matrix multiplication. The time proximity sensitivity coefficient, express Compared to The repulsion coefficient of the node message propagation, which is a key power quality parameter to be propagated.

[0075] Step S3.2.3: Based on the sequence distribution of the repulsion coefficient of the node message propagation of the key power quality parameter to be propagated and the sequence distribution of the dynamic transmission potential energy encoding vector of the node message propagation of the key power quality parameter to be propagated, perform dynamic adaptive message propagation encoding on each structured encoding vector of the key power quality parameter to be propagated in the sequence distribution of the structured encoding vector of the key power quality parameter to be propagated, so as to obtain the set of dynamic encoding vectors of key power quality parameter messages, specifically, expressed by the formula:

[0076] ;

[0077] ;

[0078] in, and This represents dot product and dot subtraction by position. Represents matrix multiplication. Represents the sigmoid activation function. Represents the linear rectified function. This represents the third weight matrix in the pre-training process. express The corresponding dynamic encoding vector of key power quality parameter messages.

[0079] The specific implementation of step S3.3 is as follows:

[0080] Step S3.3.1: Calculate the positional summation of the set of dynamic coding vectors for the key power quality parameter messages to obtain the dynamic propagation aggregate coding vector for the key power quality parameter messages;

[0081] Step S3.3.2: Perform weighted fusion of the dynamic propagation aggregation encoding vector of the key power quality parameter message and the current key power quality parameter structured encoding vector to obtain the key power quality parameter temporal context encoding vector.

[0082] Furthermore, step S3.3.2 is specifically implemented as follows:

[0083] The dynamic-static feature interaction sensing is performed on the dynamic propagation aggregated encoding vector of the key power quality parameter message and the current key power quality parameter structured encoding vector to obtain the fusion weight coefficient. Specifically, it is expressed by the formula:

[0084] ;

[0085] in, This represents the fourth weight matrix in the pre-training process. Represents the normalized exponential function, through Function operations can produce binary vectors. Indicates vector concatenation. Indicates the fusion weight coefficient;

[0086] The fusion weight coefficient is used as the weight of the current key power quality parameter structured encoding vector, and the difference between the fusion weight coefficient and the fusion weight coefficient (i.e., 1 minus the fusion weight coefficient) is used as the weight of the key power quality parameter message dynamic propagation aggregation encoding vector. The key power quality parameter message dynamic propagation aggregation encoding vector and the current key power quality parameter structured encoding vector are weighted and fused to obtain the key parameter time-series context encoding vector. Specifically, this is expressed by the formula:

[0087] ;

[0088] in, This represents the timing context encoding vector for key power quality parameters.

[0089] Meanwhile, for the set of dynamic encoding vectors of key power quality parameter messages, due to the adaptability of effect propagation under nonlinear representation, each dynamic encoding vector of key power quality parameter messages essentially constitutes a decoupled vector local propagation grid. Therefore, when aggregating each local propagation grid, in addition to vector summation, a local propagation grid propagation path correction that conforms to the sequence propagation characteristics can also be introduced.

[0090] Specifically, a vector product state is introduced to simulate the auxiliary propagation path representation of local propagation lattice points, thereby using the product structure to reflect the propagation behavior of local lattice points in the sequence path, so that the synergistic effect can emerge naturally. Specifically, it is expressed by the formula:

[0091] ;

[0092] This is used to optimize the timing context encoding vector of the key power quality parameters, specifically expressed by the formula:

[0093] ;

[0094] This strengthens the propagation synergy effect by introducing auxiliary propagation paths with local propagation grids, thereby improving the aggregation effect of the dynamic time-series information of the key power quality parameter message dynamic encoding vector at each time point.

[0095] Preferably, in step S4, the key power quality parameter time-series context encoding vector is subjected to feature renormalization based on power quality parameter node collaboration to obtain an optimized key power quality parameter time-series context encoding vector. This optimized vector is then input into a decoder-based voltage waveform reconstruction module to obtain the reconstructed voltage waveform data. Specifically, during the training phase, the decoder is trained using power quality data under normal conditions (i.e., key power quality parameter time-series context encoding vectors extracted from voltage waveform data without significant disturbances or abnormal precursors) to learn the latent representation of normal power quality conditions. By optimizing the decoder parameters, the decoder learns the inherent, typical, and regular feature patterns of normal power quality data, thereby accurately reconstructing the original voltage waveform data from the key power quality parameter time-series context encoding vector under normal voltage conditions. During the inference phase, when the input is a key power quality parameter time-series context encoding vector under normal conditions, the decoder can effectively utilize its learned prior knowledge to generate normal voltage waveform data that corresponds to the input and is as close as possible to the original. When the input is a time-series context encoding vector of key power quality parameters under significantly abnormal conditions, the decoder attempts to reconstruct the voltage waveform based on the learned latent representation of the normal state. However, due to the differences between the abnormal and normal states in the feature space, the reconstructed voltage waveform will deviate from the actually observed waveform. This deviation can be used as an indicator of anomaly detection; the larger the deviation, the higher the degree of anomaly in the current voltage waveform data.

[0096] Preferably, in step S5, the reconstruction error is the root mean square error between the voltage waveform data and the reconstructed voltage waveform data. Based on the comparison between the reconstruction error and a preset threshold, it is determined whether there are any abnormal precursors. It should be understood that since the reconstruction error directly reflects the degree of deviation between the input voltage waveform data and the normal pattern learned by the decoding model, this invention uses it as an anomaly criterion. By calculating the root mean square error between the voltage waveform data and the reconstructed voltage waveform data, the overall deviation between the actual observed waveform and the reconstructed waveform is quantified, serving as a quantitative basis for power quality anomaly detection.

[0097] When the reconstruction error exceeds a preset threshold, it indicates a significant anomaly in the current voltage waveform data, potentially signifying a fault or instability in the power grid. The system immediately issues an alarm, indicating a possible power quality problem, thus enabling real-time monitoring and early warning of power quality. This method achieves automatic and sensitive detection of precursors to power quality anomalies, fulfilling the goal of early warning and providing strong support for preventative maintenance to ensure stable power grid operation.

[0098] like Figure 1 As shown, the present invention also discloses a power quality monitoring device optimized based on an AI engine, comprising:

[0099] The voltage waveform data acquisition module is used to acquire voltage waveform data collected by the monitoring equipment.

[0100] A voltage waveform short-time window segmentation module is used to segment the voltage waveform data into short-time windows to obtain a sequence distribution of the voltage waveform short-time window data;

[0101] The power quality feature extraction module is used to extract the power quality features of each voltage waveform short-time window data in the sequence distribution of the voltage waveform short-time window data, so as to obtain the sequence distribution of the structured encoding vector of key power quality parameters;

[0102] The power quality parameter context encoding module is used to perform context encoding based on temporal adversarial repulsion guidance on the sequence distribution of the structured encoding vector of the key power quality parameters to obtain the temporal context encoding vector of the key power quality parameters.

[0103] The reconstructed voltage waveform data generation module is used to generate reconstructed voltage waveform data based on the timing context encoding vector of the key power quality parameters.

[0104] An abnormal precursor monitoring module is used to determine whether there are abnormal precursors based on the reconstruction error between the voltage waveform data and the reconstructed voltage waveform data.

[0105] Compared with the prior art, the present invention proposes the following innovations:

[0106] 1) The voltage waveform data is segmented into short-time windows, and multivariate key power quality parameters are extracted from the voltage waveform data of each window to construct a time-series distribution of multidimensional power quality parameters to characterize the evolution of the power state. Then, a deep learning algorithm is introduced to extract features and model the temporal context of the key power quality parameters in each short-time window, learning the temporal evolution law of the power quality state. Based on this, the prior knowledge learned by the decoder under normal power quality data is used to reconstruct the voltage waveform data. Furthermore, the reconstruction error between the reconstructed and original voltage waveform data is used to identify abnormal precursors. This method utilizes the learning ability of AI models on the waveform characteristics of normal power state to achieve early warning of power quality disturbance events, facilitating preventative maintenance and proactive intervention.

[0107] 2) Since voltage waveform data is essentially time-series data, it contains rich information on the historical evolution of voltage waveforms and clues for predicting future trends. Therefore, in order to fully explore the time-series dependencies and abnormal evolution patterns in voltage waveform data and enhance the model's sensitivity to key abnormal events, this invention proposes a context encoding method based on time-series adversarial repulsion guidance. This method simulates the concepts of potential energy and repulsion in physics, treating the structured encoding vectors of key power quality parameters at historical time nodes as "sources" with different influences on the current state. Based on the distance (time interval) between historical nodes and the current node and their correlation with power quality state, the propagation potential energy and repulsion of the structured encoding vectors of key power quality parameters at each historical node are calculated. This is used to dynamically modulate the vectors, enhancing the model's sensitivity to abnormal patterns. By aggregating the context information of global nodes, a time-series context encoding vector of key power quality parameters that integrates historical evolution patterns and clues for predicting future trends is obtained, providing a more refined feature representation for subsequent voltage waveform data reconstruction.

[0108] It is worth mentioning that the technical features such as the PQ monitoring equipment involved in this patent application should be regarded as prior art. The specific structure, working principle, and possible control methods and spatial arrangement of these technical features can be adopted using conventional choices in the field, and should not be regarded as the inventive point of this patent. This patent will not be further elaborated in detail.

[0109] For those skilled in the art, modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the protection scope of this invention.

Claims

1. A power quality monitoring method based on AI engine optimization, characterized by, The method comprises the following steps: Step S1: acquiring voltage waveform data collected by a monitoring device, and performing short-time window segmentation on the voltage waveform data to obtain a sequence distribution of voltage waveform short-time window data; Step S2: extracting an electric energy quality parameter sequence of each voltage waveform short-time window data in the sequence distribution of the voltage waveform short-time window data to obtain a sequence distribution of a key electric energy quality parameter structured coding vector; Step S3: performing context coding based on time sequence antipodal repulsion guidance on the sequence distribution of the key electric energy quality parameter structured coding vector to obtain a key electric energy quality parameter time sequence context coding vector; Step S3 is specifically implemented as the following steps: Step S3.1: extracting a current key electric energy quality parameter structured coding vector from the sequence distribution of the key electric energy quality parameter structured coding vector, and defining other key electric energy quality parameter structured coding vectors in the sequence distribution of the key electric energy quality parameter structured coding vector as to-be-propagated key electric energy quality parameter structured coding vectors to obtain a sequence distribution of to-be-propagated key electric energy quality parameter structured coding vectors; Step S3.2: performing feature propagation modulation on each to-be-propagated key electric energy quality parameter structured coding vector in the sequence distribution of the to-be-propagated key electric energy quality parameter structured coding vectors based on the time sequence correlation of each to-be-propagated key electric energy quality parameter structured coding vector in the sequence distribution of the to-be-propagated key electric energy quality parameter structured coding vectors relative to the current key electric energy quality parameter structured coding vector to obtain a set of key electric energy quality parameter message dynamic coding vectors; Step S3.3: fusing the set of key electric energy quality parameter message dynamic coding vectors and the current key electric energy quality parameter structured coding vector to obtain the key electric energy quality parameter time sequence context coding vector; Step S3.2 is specifically implemented as: Step S3.2.1: calculating a dynamic transmission potential of each to-be-propagated key electric energy quality parameter structured coding vector in the sequence distribution of the to-be-propagated key electric energy quality parameter structured coding vectors relative to the current key electric energy quality parameter structured coding vector to obtain a sequence distribution of to-be-propagated key electric energy quality parameter node dynamic transmission potential coding vectors; Step S3.2.2: calculating a message propagation repulsion force coefficient of each to-be-propagated key electric energy quality parameter structured coding vector in the sequence distribution of the to-be-propagated key electric energy quality parameter structured coding vectors relative to the current key electric energy quality parameter structured coding vector to obtain a sequence distribution of to-be-propagated key electric energy quality parameter node message propagation repulsion force coefficients; Step S3.2.3: performing dynamic adaptive message propagation coding on each of the sequence distribution of the key power quality parameter structured coding vectors based on the sequence distribution of repulsion force coefficients of the key power quality parameter node message to be propagated and the sequence distribution of the key power quality parameter node dynamic transmission potential energy coding vector to be propagated, to obtain a set of key power quality parameter message dynamic coding vectors; For step S3.3, the following is implemented: Step S3.3.1: calculating the position-wise sum of the set of key power quality parameter message dynamic coding vectors to obtain a key power quality parameter message dynamic propagation aggregation coding vector; Step S3.3.2: performing weighted fusion on the key power quality parameter message dynamic propagation aggregation coding vector and the current key power quality parameter structured coding vector to obtain the key power quality parameter time series context coding vector; Step S4: generating reconstructed voltage waveform data based on the key power quality parameter time series context coding vector; Step S5: determining whether there is an abnormal precursor based on the reconstruction error between the voltage waveform data and the reconstructed voltage waveform data.

2. The method for power quality monitoring based on AI engine optimization according to claim 1, characterized in that, In step S2, the power quality parameter sequence of each voltage waveform short-time window data in the sequence distribution is extracted to obtain a sequence distribution of key power quality parameter sequences; each key power quality parameter sequence in the sequence distribution of key power quality parameter sequences is structured coded using a key power quality parameter embedding coding matrix to obtain the sequence distribution of key power quality parameter structured coding vectors.

3. The method for power quality monitoring based on AI engine optimization according to claim 2, characterized in that, For step S3.3.2, the following is implemented: Performing dynamic-static feature interaction perception on the key power quality parameter message dynamic propagation aggregation coding vector and the current key power quality parameter structured coding vector to obtain a fusion weight coefficient; using the fusion weight coefficient as the weight of the current key power quality parameter structured coding vector, using the difference between one and the fusion weight coefficient as the weight of the key power quality parameter message dynamic propagation aggregation coding vector, performing weighted fusion on the key power quality parameter message dynamic propagation aggregation coding vector and the current key power quality parameter structured coding vector to obtain the key power quality parameter time series context coding vector.

4. The power quality monitoring method based on AI engine optimization according to claim 3, characterized in that, In step S4, performing feature renormalization based on power quality parameter node cooperation on the key power quality parameter time series context coding vector to obtain an optimized key power quality parameter time series context coding vector; inputting the optimized key power quality parameter time series context coding vector into a voltage waveform reconstruction module based on a decoder to obtain the reconstructed voltage waveform data.

5. The method for power quality monitoring based on AI engine optimization according to claim 4, characterized in that, In step S5, the reconstruction error is the root mean square error between the voltage waveform data and the reconstructed voltage waveform data; based on the comparison between the reconstruction error and a preset threshold, it is determined whether there is an abnormal precursor.

6. An AI engine optimization-based power quality monitoring device applied to the AI engine optimization-based power quality monitoring method of any one of claims 1-5, characterized in that, It includes: A voltage waveform data acquisition module configured to acquire voltage waveform data collected by a monitoring device; a voltage waveform short-time window segmentation module, configured to perform short-time window segmentation on the voltage waveform data to obtain a sequence distribution of voltage waveform short-time window data; an electric energy quality feature extraction module, configured to extract an electric energy quality feature of each voltage waveform short-time window data in the sequence distribution of voltage waveform short-time window data to obtain a sequence distribution of key electric energy quality parameter structured coding vectors; an electric energy quality parameter context coding module, configured to perform context coding on the sequence distribution of key electric energy quality parameter structured coding vectors based on time sequence counteracting repulsive force guidance to obtain a key electric energy quality parameter time sequence context coding vector; a reconstructed voltage waveform data generation module, configured to generate reconstructed voltage waveform data based on the key electric energy quality parameter time sequence context coding vector; an abnormal precursor monitoring module, configured to determine whether there is an abnormal precursor based on a reconstruction error between the voltage waveform data and the reconstructed voltage waveform data.

Images