A multi-voltage-grade power quality on-line monitoring cross-site comprehensive analysis system

Abstract

The application discloses a kind of multi-voltage grade electric energy quality on-line monitoring cross-station comprehensive analysis system, and relates to power system monitoring technical field.Edge perception module: set edge perception layer, for real-time acquisition voltage, current waveform data, get power station data item, while edge perception layer built-in voltage adaptive range switching module.The application is deployed by including wideband sensing terminal in each different voltage grade substation edge perception layer, for real-time acquisition voltage, current waveform data, and set regional coordination layer, regional coordination layer connects multiple edge perception layer nodes, based on power station data item carries out data collection, carries out data splitting by model creation module, creates graph neural network model, based on graph neural network model carries out disturbance tracing, and outputs global analysis result, to realize the cross-station comprehensive analysis in the multi-dimensional state of data dimension, time dimension, analysis dimension.

Description

Technical Field

[0001] This invention relates to the field of power system monitoring technology, specifically to a cross-station integrated analysis system for online monitoring of power quality at multiple voltage levels. Background Technology

[0002] A power quality analysis system is a comprehensive platform or software tool used to acquire, monitor, analyze, and evaluate the power quality status of a power supply system. Its core objective is to ensure that the power provided by the power system meets technical standards and user needs, and to reduce the impact of harmonics, transients, voltage fluctuations, and other issues on equipment and user experience. A power quality analysis system is a comprehensive platform that performs real-time or near-real-time acquisition, processing, evaluation, alarming, and visualization of quantities related to power quality in the power system, such as voltage, current, frequency, phase, waveform distortion, harmonics, transient waveforms, and power factor. Its goal is to ensure the stability, reliability, and availability of power, helping operation and maintenance, dispatching, and design personnel to quickly diagnose problems, locate their sources, assess their impact, and verify the effectiveness of improvements.

[0003] A power quality analysis method and apparatus, disclosed in patent publication number CN117520926A, includes a network training unit. The first neural network comprises an input layer, multiple convolutional layers, a channel attention module, an ASPP module, and an output layer. Through the network training unit, a classification cross-entropy loss function is selected as the target loss function L of the first neural network. The Adagrad optimizer is used to iteratively train the first neural network, setting an initial learning rate and continuously adjusting the weights w and biases b until the target loss function value is minimized. This method combines wavelet decomposition and neural network technology, eliminating the need for complex preprocessing and enabling rapid and accurate identification of various disturbance types in power signals.

[0004] In the application of the above and similar technical solutions, existing power quality analysis systems are mostly single-site or local multi-source aggregation, making it difficult to achieve unified data formats, clock synchronization, indicator definitions, and alarm strategies in cross-site and cross-voltage level scenarios. This leads to difficulties in power quality comparison, source tracing, and cross-site cascading analysis at the network-wide level. Furthermore, the lack of a unified cross-site data model and semantic standards results in inconsistencies in timing alignment, unit dimensions, feature definitions, and event labeling of data from different manufacturers' equipment, which in turn leads to deviations in power quality analysis results. Summary of the Invention

[0005] The purpose of this invention is to provide a cross-station integrated analysis system for online monitoring of power quality at multiple voltage levels, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a cross-station integrated analysis system for online monitoring of power quality at multiple voltage levels, comprising:

[0007] Edge sensing module: An edge sensing layer is set up to collect voltage and current waveform data in real time to obtain power plant data items. At the same time, the edge sensing layer has a built-in voltage adaptive range switching module to perform adaptive range adjustment based on the power plant data items to obtain range adjustment items.

[0008] Regional Collaboration Module: A regional collaboration layer is set up, which connects multiple edge sensing layer nodes. Data sets are generated based on power station data items to obtain a power station dataset. The power station dataset includes at least one power station data item. Timestamp synchronization is performed based on the power station dataset to achieve cross-station data spatiotemporal alignment and obtain a power station data alignment set.

[0009] Model creation module: Based on the power plant data alignment set, the data is split into node features and edge features. A node relationship matrix is ​​constructed based on the node features and edge features. A graph convolutional network is used for model training to create a graph neural network model.

[0010] Cloud-based control module: A cloud-based intelligent layer is set up. The cloud-based intelligent layer performs disturbance source tracing based on a graph neural network model and outputs global analysis results, thereby realizing cross-site comprehensive analysis in multiple dimensions such as data dimension, time dimension, and analysis dimension.

[0011] Furthermore, the method for obtaining the power plant data items includes:

[0012] The number of edge sensing layers is at least one, which is deployed in the target voltage level substation and includes a broadband sensing terminal. The broadband sensing terminal includes a voltage transformer unit, a signal conditioning circuit and an edge computing unit.

[0013] The target frequency response range is obtained by setting the frequency response range based on the broadband sensing terminal, and the real-time data of the voltage level substation is obtained based on the voltage transformer unit, signal conditioning circuit and edge computing unit, thus obtaining the substation data item.

[0014] Furthermore, the method for obtaining the range adjustment term includes:

[0015] Based on the voltage adaptive range switching module, at least two voltage ranges are set to obtain voltage range items. Based on the voltage range items, the corresponding impedance data are set respectively to obtain at least two impedance information items.

[0016] Based on the real-time matching data of the power plant data item and the voltage range item, the real-time voltage range item is obtained. Based on the real-time voltage range item, the corresponding impedance data is obtained, and then the target impedance information item is obtained. The range adjustment item is obtained by using the target impedance information item as a reference.

[0017] Furthermore, the method for obtaining the power plant dataset includes:

[0018] Based on the power station data items, the substations at the target voltage level are labeled to obtain the power station label items, and the power station data items and power station label items are matched.

[0019] A power grid topology matrix is ​​constructed, and electrical connection parameters between substations of the target voltage level are stored based on power station data items and power station label items, thereby obtaining the power station dataset.

[0020] Furthermore, the method for obtaining the power plant data alignment set includes:

[0021] Based on the power plant dataset, a target protocol method is defined, and cross-station clock synchronization is achieved through the target protocol method.

[0022] The target timestamp is used for timestamp synchronization. A time error threshold is set, which is a fixed time value. Based on the target timestamp, the time error is ensured to be lower than the error threshold. Then, the power plant dataset is spatiotemporally aligned to obtain the power plant data alignment set.

[0023] Furthermore, the node features include electrical energy data corresponding to the power plant data items, including voltage data and current waveform data. The method for obtaining the node features includes:

[0024] Feature extraction is performed hierarchically based on the alignment set of power plant data, including time-domain feature extraction and frequency-domain feature extraction;

[0025] Time-domain feature extraction includes fundamental parameter extraction, obtaining voltage data, current phase difference data, and power factor, and obtaining voltage sag depth values ​​based on voltage data, current phase difference data, and power factor;

[0026] Frequency domain feature extraction includes harmonic spectrum extraction. Based on power plant data items, the frequency is divided into segments to obtain at least three segmented quantization terms. Feature calculation is performed based on the segmented quantization terms to obtain the energy of each frequency band and the total harmonic distortion rate.

[0027] Based on voltage sag depth, energy of each frequency band, and total harmonic distortion, a 128-dimensional node feature vector is generated, which then transforms the data in the power plant data alignment set into deep features characterizing power quality.

[0028] Furthermore, the edge features are edge-aware layer node association data corresponding to the power plant dataset, including node length, line impedance, and attenuation information. The methods for obtaining the edge features include:

[0029] Basic parameters are obtained based on the power plant data alignment set, including dynamic and static parameters. Static parameters are extracted from the power plant data alignment set, such as line length, conductor type, unit impedance, and transformer ratio. Dynamic parameters include real-time measured load current, thereby obtaining edge line characteristics.

[0030] Furthermore, the method for creating the graph neural network model includes:

[0031] The system consists of an input layer, a core network layer, and an output layer. The input layer includes node feature matrices and edge feature matrices. The core network layer includes spatiotemporal convolutional layers with temporal and spatial dimensions. The temporal dimension is used to extract transient event time-series patterns, and the spatial dimension is used to aggregate neighborhood node information. The output layer is used for perturbation source tracing and classification, as well as propagation path reconstruction.

[0032] A node relationship matrix is ​​constructed based on node features and edge features, and a graph convolutional network is used for model training to obtain a graph neural network model.

[0033] Compared with the prior art, the beneficial effects of the present invention are:

[0034] This multi-voltage level power quality online monitoring cross-station integrated analysis system deploys an edge sensing layer containing broadband sensor terminals in substations of different voltage levels to collect voltage and current waveform data in real time. It also establishes a regional collaboration layer that connects multiple edge sensing layer nodes. Data is aggregated based on substation data items, and the data is split into node features and edge features through a model creation module. Feature extraction is performed hierarchically based on node features, including time-domain and frequency-domain feature extraction. A node relationship matrix is ​​constructed based on the node and edge features, and a graph convolutional network is used for model training to create a graph neural network model. The cloud-based intelligent layer uses the graph neural network model to trace disturbance sources and outputs global analysis results, thus achieving multi-dimensional cross-station integrated analysis across data, time, and analysis dimensions. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0036] Figure 2 This is a schematic diagram of the power plant data item acquisition process of the present invention;

[0037] Figure 3 This is a schematic diagram of the process for obtaining the range adjustment item of the present invention;

[0038] Figure 4 This is a schematic diagram illustrating the relationship between the regional collaboration layer, nodes, and labels of the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Most existing power quality analysis systems adopt a single-site or local multi-source aggregation architecture, lacking a unified data format, clock synchronization, indicator definition, and alarm strategy. This means that monitoring equipment at different sites or voltage levels may use different data encoding methods, sampling frequencies, time bases, and indicator calculation methods. For example, power quality analyzers from different manufacturers may use different harmonic calculation standards, resulting in different harmonic content analysis results for the same grid disturbance at different sites. Regarding alarm strategies, different sites may trigger alarms based on different thresholds, making it difficult to uniformly manage and analyze alarm information across the entire network. This inconsistency makes network-wide power quality comparison, source tracing, and cross-site cascading analysis extremely difficult. Power engineers need to spend a significant amount of time and effort on data conversion, calibration, and alignment before they can even begin preliminary analysis. This significantly reduces analysis efficiency and accuracy. The technical solution provided in this application, however, deploys an edge sensing layer containing broadband sensing terminals within substations of different voltage levels for real-time voltage and current waveform data acquisition. A regional collaboration layer is established, connecting multiple edge sensing layer nodes. Data is aggregated based on substation data items, and a model creation module splits the data into node features and edge features. Feature extraction is performed hierarchically based on node features, including time-domain and frequency-domain feature extraction. A node relationship matrix is ​​constructed based on node and edge features, and a graph convolutional network is used for model training to create a graph neural network model. The cloud-based intelligent layer uses this graph neural network model to trace disturbance sources and outputs global analysis results. This achieves cross-station comprehensive analysis across multiple dimensions—data, time, and analysis—such as… Figure 1 As shown, it includes an edge perception module, a regional collaboration module, a model creation module, and a cloud control module.

[0041] Edge sensing module: An edge sensing layer is set up to collect voltage and current waveform data in real time to obtain power plant data items. At the same time, the edge sensing layer has a built-in voltage adaptive range switching module to perform adaptive range adjustment based on the power plant data items to obtain range adjustment items.

[0042] It is important to note that, such as Figure 2As shown, the method for obtaining power station data items includes: at least one edge sensing layer is deployed in the target voltage level substation, including a broadband sensing terminal, which includes a voltage transformer unit, a signal conditioning circuit, and an edge computing unit; the frequency response range is set based on the broadband sensing terminal to obtain the target frequency response range item; real-time data of the voltage level substation is obtained based on the voltage transformer unit, the signal conditioning circuit, and the edge computing unit to obtain the power station data item.

[0043] Specifically, to achieve cross-station integrated analysis, multiple edge sensing layers are deployed in substations of different voltage levels. Since the voltage, frequency, and other data transmitted by substations of different voltage levels may vary, wideband sensing terminals are also installed in the edge sensing layers. The frequency response range is set. The frequency response range of traditional sensors is 0-2.5kHz, while this application uses wideband sensing terminals with a set frequency response range of 0-150kHz, which captures a wider range. Real-time data of substations of different voltage levels is obtained based on voltage transformer units, signal conditioning circuits, and edge computing units, thereby obtaining substation data items.

[0044] It is important to note that, such as Figure 3 As shown, the method for obtaining the range adjustment item includes: based on the voltage adaptive range switching module, setting at least two voltage ranges to obtain a voltage range item; setting corresponding impedance data based on the voltage range item to obtain at least two impedance information items; obtaining a real-time voltage range item based on the real-time matching data of the power station data item and the voltage range item; obtaining corresponding impedance data based on the real-time voltage range item to obtain a target impedance information item; and adjusting the range adjustment item using the target impedance information item as a reference.

[0045] Specifically, two voltage ranges are set: a low-voltage range and a high-voltage range. The low-voltage range is 0-10kV, and the high-voltage range is 10-500kV, thus obtaining the voltage range item. Based on the voltage range item, the corresponding impedance data is set, with the impedance data corresponding to the low-voltage range being 1MΩ and the impedance data corresponding to the high-voltage range being 500MΩ. At this time, the actual voltage data obtained from the power station data item is matched with the voltage range item to obtain the real-time voltage range item. Based on the real-time voltage range item, the corresponding impedance data is obtained to obtain the target impedance information item. The range adjustment item is obtained by using the target impedance information item as a reference.

[0046] Regional Collaboration Module: A regional collaboration layer is set up, which connects multiple edge sensing layer nodes. Data sets are created based on power station data items to obtain a power station dataset. The power station dataset includes at least one power station data item. Timestamp synchronization is performed based on the power station dataset to achieve cross-station data spatiotemporal alignment and obtain a power station data alignment set.

[0047] It should be noted that the method for obtaining the power plant dataset includes: assigning labels to substations of the target voltage level based on the power plant data items to obtain power plant label items, and performing data matching between the power plant data items and the power plant label items; constructing a power grid topology matrix, and storing electrical connection parameters between substations of the target voltage level based on the power plant data items and the power plant label items, thereby obtaining the power plant dataset.

[0048] Specifically, the regional coordination layer connects multiple edge sensing layer nodes, which are deployed in substations of various voltage levels. Therefore, the substations of different voltage levels are first labeled. Based on the actual location of the regional coordination layer, the substations closest to the regional coordination layer are sorted by location and labeled. At the same time, based on the obtained substation data items, the substation data items are matched with the labeled substations to construct a power grid topology matrix. The electrical connection parameters between substations are stored and recorded to obtain the substation dataset.

[0049] In the specific implementation process, such as Figure 4 As shown, five edge sensing layer nodes, Nodes 1-5, are connected through a regional collaboration layer. Nodes 1-5 are installed in substations of different levels (Level 1-Level 5), respectively. The regional collaboration layer is located at point A. Substations of different levels (Level 1-Level 5) are located near point A, with Level 1 substations 10km, Level 2 substations 50km, Level 3 substations 100km, Level 4 substations 30km, and Level 5 substations 80km away. Nodes 1-5 obtain information about the edge sensing layer. The voltage and current waveform data of Level 1 to Level 5 are collected. At this time, substations of Level 1 to Level 5 are labeled and assigned numbers in the order of their distance from location A. Therefore, the substations of Level 1 to Level 5 are labeled a, c, e, b, and d, respectively. The substation data items obtained by nodes 1 to 5 are then matched with labels a, c, e, b, and d to construct a power grid topology matrix. The electrical connection parameters between substations are stored and recorded to obtain the substation dataset.

[0050] It should be noted that the method for obtaining the power plant data alignment set includes: based on the power plant dataset, setting a target protocol method, and achieving cross-station clock synchronization through the target protocol method; using a target timestamp for timestamp synchronization, setting a time error threshold, which is a fixed time value, and ensuring that the time error is lower than the error threshold based on the target timestamp, thereby performing spatiotemporal alignment on the power plant dataset to obtain the power plant data alignment set.

[0051] Specifically, the target protocol includes the IEEE 1588v3 protocol, which is used to achieve cross-site clock synchronization. The target timestamp used includes the FPGA hardware timestamp, which is used for timestamp synchronization. The set time error threshold is 1μs. Based on the time error threshold, the time error is limited to not exceeding the time error threshold during the timestamp alignment of the power plant dataset, thereby obtaining the power plant data alignment set.

[0052] Model creation module: Based on the power plant data alignment set, the data is split into node features and edge features. A node relationship matrix is ​​constructed based on the node features and edge features. A graph convolutional network is used for model training to create a graph neural network model.

[0053] It is important to note that node features include power data corresponding to power plant data items, including voltage data and current waveform data. The methods for obtaining node features include: hierarchical feature extraction based on the power plant data alignment set, including time-domain feature extraction and frequency-domain feature extraction; time-domain feature extraction includes fundamental parameter extraction, obtaining voltage data, current phase difference data, and power factor, and obtaining voltage sag depth values ​​based on voltage data, current phase difference data, and power factor; frequency-domain feature extraction includes harmonic spectrum extraction, dividing the frequency into segments based on the power plant data items to obtain at least three segmented quantization terms, performing feature calculations based on the segmented quantization terms to obtain the energy of each frequency band and the total harmonic distortion rate; and generating a 128-dimensional node feature vector based on the voltage sag depth values, the energy of each frequency band, and the total harmonic distortion rate, thereby transforming the data in the power plant data alignment set into deep features characterizing power quality.

[0054] Specifically, the raw data first needs to be preprocessed, including noise reduction, using adaptive wavelet threshold filtering to eliminate electromagnetic interference. Then, time-domain and frequency-domain feature extraction are performed. In the time-domain feature extraction stage, the fundamental parameters include the effective values ​​of the three-phase voltages, the current phase difference, and the power factor. The voltage sag depth is obtained based on these parameters, with time-domain features accounting for 40% of the vector dimension. Next, frequency-domain feature extraction is performed, including harmonic spectrum extraction. Based on the power plant data items, the frequency is segmented to obtain three segmented quantities. The quantization terms are defined as power frequency harmonics, intermediate frequency harmonics, and high frequency harmonics, with power frequency harmonics ranging from 0-2kHz, intermediate frequency harmonics from 2-30kHz, and high frequency harmonics from 30-150kHz. Feature calculations are performed based on the segmented quantization terms to obtain the energy of each frequency band and the total harmonic distortion rate, with frequency domain features accounting for 60% of the vector dimension. A 128-dimensional node feature vector is generated based on the voltage sag depth, the energy of each frequency band, and the total harmonic distortion rate, with an update frequency of 100ms / time, matching the power grid frequency cycle. This transforms the power plant data into a deep feature characterizing power quality.

[0055] It is important to note that the edge features are associated data of the edge sensing layer nodes corresponding to the power plant dataset, including node length, line impedance, and attenuation information. The methods for obtaining edge features include: obtaining basic parameters based on the power plant data alignment set, including dynamic parameters and static parameters. The static parameters are extracted from the power plant data alignment set, including line length, conductor type, unit impedance, and transformer ratio. The dynamic parameters include real-time measured load current, thereby obtaining the edge features.

[0056] Specifically, edge features are used to quantify the physical transmission characteristics of the connecting lines between nodes. In the process of acquiring basic parameters, static parameters are extracted from the power grid topology database, including line length, conductor type, unit impedance, and transformer ratio. Dynamic parameters are obtained by real-time measurement of load current. Based on the dynamic and static parameters, feature vectors are generated and a 64-dimensional edge feature vector is output.

[0057] It's important to note that the method for creating a graph neural network model includes: defining an input layer, a core network layer, and an output layer. The input layer includes node feature matrices and edge feature matrices. The core network layer includes spatiotemporal convolutional layers, defining temporal and spatial dimensions. The temporal dimension is used to extract transient event time-series patterns, and the spatial dimension is used to aggregate neighborhood node information. The output layer is used for perturbation source tracing and classification, as well as propagation path reconstruction. A node relationship matrix is ​​constructed based on node features and edge features, and a graph convolutional network is used for model training to obtain the graph neural network model.

[0058] Specifically, the input layer includes a node feature matrix and an edge feature matrix. The node feature matrix includes the number of nodes, and the edge feature matrix includes the number of edges. The spatiotemporal convolutional layer has a temporal dimension and a spatial dimension. The temporal dimension is used to extract transient event time-series patterns and uses a 1D convolutional kernel with a width of 10 cycles. The spatial dimension is used to aggregate neighborhood node information and uses a gated graph convolution. The output layer is used for perturbation source tracing and classification and propagation path reconstruction. The perturbation source tracing outputs the node probability distribution, and the propagation path reconstruction outputs the edge activation probability. A node relationship matrix is ​​constructed based on node features and edge features. A graph convolutional network is used for model training to obtain a graph neural network model.

[0059] Cloud-based control module: A cloud-based intelligent layer is set up. The cloud-based intelligent layer performs disturbance source tracing based on a graph neural network model and outputs global analysis results, thereby realizing cross-site comprehensive analysis in multiple dimensions such as data dimension, time dimension, and analysis dimension.

[0060] It is important to note that the cloud-based intelligent layer integrates and trains data based on the output of the graph neural network model and the data obtained from the edge perception layers set up in substations of different voltage levels. This enables disturbance tracing and propagation path reconstruction, and then outputs global analysis results. In this way, cross-station comprehensive analysis is achieved in multiple dimensions, including data, time, and analysis.

[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended embodiments and their equivalents.

Claims

1. A multi-voltage level power quality online monitoring cross-station integrated analysis system, comprising: Edge sensing module: An edge sensing layer is set up to collect voltage and current waveform data in real time to obtain power plant data items. At the same time, the edge sensing layer has a built-in voltage adaptive range switching module to perform adaptive range adjustment based on the power plant data items to obtain range adjustment items. Its characteristic is that it further includes: Regional Collaboration Module: A regional collaboration layer is set up, which connects multiple edge sensing layer nodes. Data sets are generated based on power station data items to obtain a power station dataset. The power station dataset includes at least one power station data item. Timestamp synchronization is performed based on the power station dataset to achieve cross-station data spatiotemporal alignment and obtain a power station data alignment set. Model creation module: Based on the power plant data alignment set, the data is split into node features and edge features. A node relationship matrix is ​​constructed based on the node features and edge features. A graph convolutional network is used for model training to create a graph neural network model. Cloud-based control module: A cloud-based intelligent layer is set up. The cloud-based intelligent layer performs disturbance source tracing based on a graph neural network model and outputs global analysis results, thereby realizing cross-site comprehensive analysis in multiple dimensions such as data dimension, time dimension, and analysis dimension.

2. The cross-station integrated analysis system for online monitoring of power quality at multiple voltage levels according to claim 1, characterized in that: The methods for obtaining the power plant data items include: The number of edge sensing layers is at least one, which is deployed in the target voltage level substation and includes a broadband sensing terminal. The broadband sensing terminal includes a voltage transformer unit, a signal conditioning circuit and an edge computing unit. The target frequency response range is obtained by setting the frequency response range based on the broadband sensing terminal, and the real-time data of the voltage level substation is obtained based on the voltage transformer unit, signal conditioning circuit and edge computing unit, thus obtaining the substation data item.

3. The cross-station integrated analysis system for online monitoring of power quality at multiple voltage levels according to claim 1, characterized in that: The method for obtaining the range adjustment term includes: Based on the voltage adaptive range switching module, at least two voltage ranges are set to obtain voltage range items. Based on the voltage range items, the corresponding impedance data are set respectively to obtain at least two impedance information items. Based on the real-time matching data of the power plant data item and the voltage range item, the real-time voltage range item is obtained. Based on the real-time voltage range item, the corresponding impedance data is obtained, and then the target impedance information item is obtained. The range adjustment item is obtained by using the target impedance information item as a reference.

4. The cross-station integrated analysis system for online monitoring of power quality at multiple voltage levels according to claim 2, characterized in that: The methods for obtaining the power plant dataset include: Based on the power station data items, the substations at the target voltage level are labeled to obtain the power station label items, and the power station data items and power station label items are matched. A power grid topology matrix is ​​constructed, and electrical connection parameters between substations of the target voltage level are stored based on power station data items and power station label items, thereby obtaining the power station dataset.

5. The cross-station integrated analysis system for online monitoring of power quality at multiple voltage levels according to claim 1, characterized in that: The method for obtaining the power plant data alignment set includes: Based on the power plant dataset, a target protocol method is defined, and cross-station clock synchronization is achieved through the target protocol method. The target timestamp is used for timestamp synchronization. A time error threshold is set, which is a fixed time value. Based on the target timestamp, the time error is ensured to be lower than the error threshold. Then, the power plant dataset is spatiotemporally aligned to obtain the power plant data alignment set.

6. The cross-station integrated analysis system for online monitoring of power quality at multiple voltage levels according to claim 1, characterized in that: The node features include electrical energy data corresponding to the power plant data items, including voltage data and current waveform data. The methods for obtaining node features include: Feature extraction is performed hierarchically based on the alignment set of power plant data, including time-domain feature extraction and frequency-domain feature extraction; Time-domain feature extraction includes fundamental parameter extraction, obtaining voltage data, current phase difference data, and power factor, and obtaining voltage sag depth values ​​based on voltage data, current phase difference data, and power factor; Frequency domain feature extraction includes harmonic spectrum extraction. Based on power plant data items, the frequency is divided into segments to obtain at least three segmented quantization terms. Feature calculation is performed based on the segmented quantization terms to obtain the energy of each frequency band and the total harmonic distortion rate. Based on voltage sag depth, energy of each frequency band, and total harmonic distortion, a 128-dimensional node feature vector is generated, which then transforms the data in the power plant data alignment set into deep features characterizing power quality.

7. The cross-station integrated analysis system for online monitoring of power quality at multiple voltage levels according to claim 1, characterized in that: The edge features are associated data of edge sensing layer nodes corresponding to the power station dataset, including node length, line impedance, and attenuation information. The methods for obtaining the edge features include: Basic parameters are obtained based on the power plant data alignment set, including dynamic and static parameters. Static parameters are extracted from the power plant data alignment set, such as line length, conductor type, unit impedance, and transformer ratio. Dynamic parameters include real-time measured load current, thereby obtaining edge line characteristics.

8. The cross-station integrated analysis system for online monitoring of power quality at multiple voltage levels according to claim 1, characterized in that: The method for creating the graph neural network model includes: The system consists of an input layer, a core network layer, and an output layer. The input layer includes node feature matrices and edge feature matrices. The core network layer includes a spatiotemporal convolutional layer with temporal and spatial dimensions. The temporal dimension is used to extract transient event time-series patterns, and the spatial dimension is used to aggregate neighborhood node information. The output layer is used for perturbation source tracing, classification, and propagation path reconstruction. A node relationship matrix is ​​constructed based on node features and edge features, and a graph convolutional network is used for model training to obtain a graph neural network model.

Images