Voltage sensor multi-parameter real-time monitoring method and device in Internet of Things environment

Abstract

The invention relates to the technical field of power grid monitoring, and discloses a voltage sensor multi-parameter real-time monitoring method and device in an Internet of Things environment. According to the method, real-time voltage fluctuation data of a target area is acquired by deploying a multi-channel voltage sensor array and is transmitted to an edge computing node; dynamic feature extraction is executed in the edge nodes, and a multi-dimensional feature matrix containing voltage fluctuation ratio, harmonic distortion and phase deviation features is generated; identifying voltage sag, overvoltage and harmonic resonance event characteristics based on an anomaly detection model; spatial correlation modeling is carried out in combination with power grid topological structure parameters, the coupling strength of adjacent sensor node parameters is calculated, and a multi-parameter correlation map containing event propagation paths and influence ranges is generated; inputting the voltage into a diagnosis model to generate a power grid state diagnosis report, and dynamically adjusting the sampling frequency and filtering parameters of the voltage sensor according to the power grid state diagnosis report. According to the method, real-time monitoring and analysis of multiple parameters of the power grid can be realized, and comprehensive support is provided for power grid state evaluation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power grid monitoring, in particular to a voltage sensor multi-parameter real-time monitoring method and device in an Internet of Things environment. BACKGROUND

[0002] At present, the Internet of Things technology is gradually penetrating into the power system, and the real-time monitoring of the operation state of the power grid is facing many challenges. The traditional voltage monitoring method mainly relies on single-channel sensors, and the data dimension collected is limited, which is difficult to fully reflect the complex voltage fluctuation in the power grid. When voltage sag, overvoltage or harmonic resonance events occur in the power grid, the single-channel sensor can only capture the local voltage change and cannot realize the overall perception of the event.

[0003] In the existing monitoring system, data processing is mainly concentrated in the cloud, and the transmission of a large amount of raw data not only occupies a large amount of communication bandwidth, but also causes data processing delay, which is difficult to meet the demand of real-time monitoring. Although the introduction of edge computing has alleviated this problem to some extent, how to perform effective dynamic feature extraction at the edge node to accurately identify various voltage events is still a difficult problem to be solved.

[0004] The topology of the power grid is complex, and the voltage events in different regions often have certain spatial correlation. However, the traditional monitoring method lacks modeling analysis of this correlation, and it is difficult to determine the propagation path and influence range of the event, which makes it difficult for power grid operation and maintenance personnel to develop accurate response strategies when dealing with faults. Moreover, the sampling frequency and filtering parameters of the sensor are usually fixed and cannot be dynamically adjusted according to the real-time state of the power grid, which to some extent affects the accuracy and effectiveness of the monitoring data. SUMMARY

[0005] The purpose of the present application is to provide a voltage sensor multi-parameter real-time monitoring method and device in an Internet of Things environment to solve the problems raised in the background.

[0006] To achieve the above purpose, the present application provides a voltage sensor multi-parameter real-time monitoring method in an Internet of Things environment, which comprises:

[0007] A multi-channel voltage sensor array deployed at the nodes of the power grid collects real-time voltage fluctuation data of the target area, and transmits the collected raw voltage data stream to the edge computing node;

[0008] Performing dynamic feature extraction operation on the raw voltage data stream in the edge computing node to generate a multi-dimensional feature matrix containing voltage fluctuation rate feature, harmonic distortion feature and phase shift feature;

[0009] analyze a change trend of each type of feature vector in the multi-dimensional feature matrix based on a preset anomaly detection model, and identify a voltage sag event feature, an overvoltage event feature, and a harmonic resonance event feature;

[0010] model spatial correlation of the identified voltage sag event feature, overvoltage event feature, and harmonic resonance event feature according to a power grid topology structure parameter, calculate a parameter coupling strength between adjacent sensor nodes, and generate a multi-parameter correlation graph containing an event propagation path and an influence range;

[0011] input the multi-parameter correlation graph into a diagnosis model to generate a power grid state diagnosis report, and dynamically adjust a sampling frequency and a filtering parameter of a voltage sensor according to a content of the diagnosis report.

[0012] Preferably, the specific process of the dynamic feature extraction operation includes:

[0013] divide the original voltage data stream by using a sliding time window, and synchronously calculate a voltage effective value change gradient, an energy ratio of a fundamental component and a harmonic component, and an absolute value of a three-phase voltage phase difference in each time window;

[0014] compare the voltage effective value change gradient with a preset reference fluctuation threshold to generate a normalized fluctuation rate feature vector;

[0015] construct a harmonic distortion feature vector according to the energy ratio of the harmonic component and the absolute value of the phase difference;

[0016] splice the normalized fluctuation rate feature vector, the harmonic distortion feature vector, and the phase offset feature vector into a three-dimensional tensor structure, and generate a multi-dimensional feature matrix by using a convolution kernel to compress a channel dimension.

[0017] Preferably, the anomaly detection model is constructed in the following manner:

[0018] extract a sag event sample, an overvoltage event sample, and a harmonic resonance event sample labeled in historical voltage data;

[0019] calculate a feature distribution interval of each type of event sample on the multi-dimensional feature matrix, and establish a sag event feature distribution boundary, an overvoltage event feature distribution boundary, and a harmonic resonance event feature distribution boundary;

[0020] when more than a preset proportion of vector points in the real-time generated multi-dimensional feature matrix break through the feature distribution boundary, activate an event feature identification flag of a corresponding type.

[0021] Preferably, the operation steps of the spatial correlation modeling include:

[0022] obtain an electrical connection relationship parameter of each sensor node in a power grid topology structure;

[0023] The time delay correlation coefficient of the event feature identification mark between adjacent nodes is calculated, and the time delay correlation coefficient is weighted and fused with the electrical connection relationship parameter to generate a spatial correlation weight;

[0024] A directed connection edge between nodes is constructed according to the spatial correlation weight, and an initial correlation graph containing the event feature propagation direction and weight is formed.

[0025] Preferably, the generation mode of the multi-parameter correlation graph is:

[0026] The power grid load distribution parameter and the environmental temperature and humidity parameter are superimposed in the initial correlation graph;

[0027] The graph neural network is used to aggregate the node load distribution parameter and the environmental temperature and humidity parameter to generate a load-environment coupling feature vector;

[0028] The load-environment coupling feature vector and the spatial correlation weight are weighted by attention, and the weight coefficient of the directed connection edge is updated;

[0029] According to the updated weight coefficient, the event feature propagation path is probabilistically reordered to generate a multi-parameter correlation graph with weight annotation.

[0030] Preferably, the workflow of the diagnosis model includes:

[0031] The topological structure of the event feature propagation path in the multi-parameter correlation graph is analyzed, and the key path node sequence is extracted;

[0032] The voltage recovery time parameter, the harmonic attenuation rate parameter and the phase synchronization deviation parameter of the key path node sequence are calculated;

[0033] The above parameters are input into the pre-trained decision tree classifier, and the power grid insulation aging risk level, the capacitor compensation failure probability and the transformer overload early warning identifier are output.

[0034] Preferably, the way to dynamically adjust the sampling frequency and the filtering parameter of the voltage sensor is:

[0035] When the diagnosis report contains the transformer overload early warning identifier, the sampling frequency of the associated sensor node is increased to a preset highest threshold;

[0036] When the diagnosis report shows that the harmonic attenuation rate is lower than the reference value, the adaptive band-stop filter is activated and the center frequency parameter thereof is configured.

[0037] Preferably, the training process of the diagnosis model includes:

[0038] The historical multi-parameter correlation graph and its corresponding power grid fault diagnosis label are collected;

[0039] Constructing a bidirectional long short-term memory network to encode the topological features of the multi-parameter correlation graph;

[0040] Generating a prediction diagnosis report by decoding the topological features through a multilayer perceptron;

[0041] Calculating the cross-entropy loss value of the prediction diagnosis report and the power grid fault diagnosis label, and updating the network parameters by backpropagation.

[0042] Preferably, the method further comprises:

[0043] When the type of the activated event feature identification flag in the real-time generated multi-dimensional feature matrix exceeds the historical sample range, performing a feature distribution boundary dynamic expansion operation:

[0044] Extracting an abnormal feature vector set that breaks through the feature distribution boundary;

[0045] Taking the convex hull boundary of the abnormal feature vector set as a reference, expanding the transient voltage event feature distribution boundary, the overvoltage event feature distribution boundary, and the harmonic resonance event feature distribution boundary by a preset proportion;

[0046] Updating the feature distribution boundary parameters of the anomaly detection model and synchronizing them to the edge computing node.

[0047] Preferably, the present application further comprises a voltage sensor multi-parameter real-time monitoring device in an Internet of Things environment, which is used to realize the voltage sensor multi-parameter real-time monitoring method in an Internet of Things environment as described above, and the device comprises:

[0048] A multi-channel voltage data acquisition unit is used to acquire raw voltage data streams through a sensor array and transmit them to an edge computing node;

[0049] A dynamic feature extraction unit is used to generate a multi-dimensional feature matrix containing voltage fluctuation rate features, harmonic distortion features, and phase shift features in the edge computing node;

[0050] An abnormal event identification unit is used to identify voltage sag event features, overvoltage event features, and harmonic resonance event features based on an anomaly detection model;

[0051] A spatial correlation modeling unit is used to generate a multi-parameter correlation graph containing event propagation paths and impact ranges according to power grid topology structure parameters;

[0052] A power grid diagnosis unit is used to output a power grid state diagnosis report according to the multi-parameter correlation graph;

[0053] A parameter adaptive adjustment unit is used to dynamically control the sampling frequency and filtering parameters of the voltage sensor according to the diagnosis report.

[0054] Compared with the prior art, the present application has the following advantages:

[0055] By deploying a multi-channel voltage sensor array to collect real-time voltage fluctuation data, the voltage information of the power grid can be obtained from multiple dimensions, which can more comprehensively reflect the running state of the power grid compared with a single-channel sensor. The raw data collected is transmitted to an edge computing node for processing, which reduces the amount of data transmitted to the cloud, reduces the occupation of communication bandwidth, and also shortens the data processing time, which helps to timely discover abnormal situations in the power grid.

[0056] The dynamic feature extraction operation is performed in the edge computing node to generate a multi-dimensional feature matrix containing multiple features, making the description of voltage features more rich, providing a more comprehensive basis for subsequent identification of various voltage events. Based on the preset abnormality detection model, the multi-dimensional feature matrix can accurately identify event features such as voltage sag, overvoltage and harmonic resonance, etc., so that various potential voltage problems in the power grid can be revealed.

[0057] The spatial correlation modeling is combined with the power grid topology structure parameters to calculate the parameter coupling strength between adjacent sensor nodes, and a multi-parameter correlation atlas is generated, which can clearly present the propagation path and influence range of the event, and facilitate the operation and maintenance personnel to understand the overall situation of the event. The multi-parameter correlation atlas is input into the diagnosis model to generate a power grid state diagnosis report, and the sampling frequency and filtering parameters of the sensor are dynamically adjusted according to the report, so that the working state of the sensor can be adapted to the real-time situation of the power grid, thereby better capturing voltage changes in the power grid and improving the pertinence and adaptability of monitoring.

[0058] This method forms a complete closed-loop monitoring process from data collection, processing, feature extraction, event identification to state diagnosis and parameter adjustment, and each link cooperates with each other, so that the monitoring of the power grid is more accurate and efficient, and can better adapt to the complex and variable power grid operating environment. Through accurate identification and analysis of various voltage events and dynamic adjustment of sensor parameters, the power grid monitoring system can maintain good performance under different operating conditions, which helps the operation and maintenance personnel to more timely and comprehensively master the status of the power grid, so as to take appropriate measures to ensure the stable operation of the power grid. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 The working principle diagram of the voltage sensor multi-parameter real-time monitoring method in the Internet of Things environment described in the application;

[0060] Figure 2 The flowchart of the dynamic feature extraction operation;

[0061] Figure 3 The flowchart of the spatial correlation modeling;

[0062] Figure 4 The flowchart of the multi-parameter correlation atlas generation;

[0063] Figure 5 Flow chart for training of diagnostic model. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0065] Please refer to Figure 1 The present application provides a voltage sensor multi-parameter real-time monitoring method in an Internet of Things environment, which comprises:

[0066] In the Internet of Things environment, the voltage sensor multi-parameter real-time monitoring method collects real-time voltage fluctuation data of a target area through a multi-channel voltage sensor array deployed at a power grid node. The sensor array acquires three-phase voltage waveforms in a synchronous sampling mode, with a sampling accuracy of not less than 0.5 level and a sampling frequency range set to 1 kHz to 10 kHz adjustable. The original voltage data stream is transmitted to an edge computing node through an industrial Ethernet or a 5G private network, with a transmission delay controlled within 10 ms. The edge computing node uses an FPGA chip to realize dynamic feature extraction, and performs sliding window processing on the input original voltage data stream, with a window length of 10 power frequency cycles and a sliding step of 1 cycle. The voltage effective value change gradient, harmonic component energy ratio and phase difference absolute value are calculated synchronously in each window, and a three-dimensional tensor containing voltage fluctuation rate features, harmonic distortion features and phase shift features is generated. The tensor channel dimension is compressed to 32 dimensions through a 1x1 convolution kernel, and a multi-dimensional feature matrix is output.

[0067] The anomaly detection model is deployed in the GPU module of the edge computing node, and uses a feature space analysis method based on Mahalanobis distance. The model loads the feature distribution boundary parameters of pre-trained temporary drop events, overvoltage events and harmonic resonance events, and monitors the distribution state of vector points in the multi-dimensional feature matrix in real time. When more than 15% of the vector points break through the feature distribution boundary of a certain type of event, the corresponding type of event feature recognition flag is triggered. The power grid topology structure parameters are imported from the SCADA system, including node impedance matrix, electrical connection relationship diagram and load distribution data. The spatial correlation modeling module calculates the time delay correlation coefficient of event features between adjacent nodes, generates a 0 to 1 normalized spatial correlation weight combined with the impedance matrix, and constructs an initial correlation graph with a weighted directed graph structure. The diagnostic model analyzes the key propagation path in the graph, outputs a power grid state diagnosis report containing insulation aging risk level and capacitor compensation failure probability, and issues a sampling frequency adjustment instruction to the sensor terminal through the Modbus protocol.

[0068] Embodiment 1: refer to Figure 2 The specific implementation process of the dynamic feature extraction operation involves multiple steps, from the processing of the original voltage data stream to the generation of the multi-dimensional feature matrix, each link needs to accurately configure parameters and use specific algorithms to complete. After the original voltage data stream enters the edge computing node, data preprocessing is first performed, including power frequency synchronous sampling and anti-aliasing filtering. Power frequency synchronous sampling ensures that data acquisition is strictly aligned with the power frequency, avoiding spectrum leakage problems. Anti-aliasing filtering uses a Butterworth low-pass filter with a cutoff frequency set to 40% of the sampling frequency to eliminate high-frequency noise interference.

[0069] The sliding time window is divided using a combination of fixed length and step length. The window length is set to 200 milliseconds, covering 10 complete 50Hz power frequency cycles, ensuring that each window contains sufficient waveform information to reflect the voltage dynamic characteristics. The sliding step is set to 20 milliseconds, i.e. each window slides forward by one power frequency cycle, ensuring the continuity of feature extraction. In each window, the system synchronously calculates three key indicators: voltage effective value change gradient, harmonic component energy ratio, and three-phase voltage phase difference absolute value.

[0070] The calculation of the voltage effective value change gradient is based on instantaneous voltage fluctuation analysis. First, the voltage waveform of each sampling point in the window is calculated to form a time series. Then, the first-order difference method is used to process the sequence to obtain the instantaneous fluctuation rate curve. Finally, the slope of the curve is fitted by the least squares method, which is taken as the fluctuation rate characteristic value of the window. This process can capture the dynamic characteristics of voltage transient increase, transient decrease and other rapid change events.

[0071] The calculation of the harmonic component energy ratio relies on Fast Fourier Transform (FFT) for frequency spectrum analysis. The system performs FFT transform on the voltage waveform in the window to extract the amplitude of the fundamental component and the 3rd, 5th, 7th, etc. main harmonic components. The total harmonic distortion rate is calculated by the ratio of the sum of the squares of the harmonic component amplitudes to the square of the fundamental component amplitude, reflecting the distortion degree of the voltage waveform. In addition, the system also calculates the proportion of odd harmonics and even harmonics in the total harmonics, respectively, to distinguish different types of harmonic pollution sources.

[0072] The calculation of the three-phase voltage phase difference is achieved through the phase-locked loop technology. The system detects the zero-crossing points of A, B, and C three-phase voltage waveforms and records their time stamps. By comparing the time difference of the three-phase zero-crossing points, the absolute value of the phase difference between two phases is calculated, and the maximum value is taken as the phase offset characteristic of the window. This indicator can reflect the problems of three-phase voltage imbalance or phase synchronization anomaly.

[0073] The generation process of the normalized fluctuation rate feature vector involves comparison and linear transformation of reference values. The reference fluctuation threshold is set according to the grid operation standard as a reference value for judging whether the voltage fluctuation is abnormal. When the voltage effective value change gradient exceeds the threshold, the system performs normalization processing in proportion to map the gradient value to a specific range; if it does not exceed the threshold, the ratio of the gradient value to the threshold is directly used as the feature value. The normalized fluctuation rate feature vector can intuitively reflect the relative strength of voltage fluctuation.

[0074] The construction of the harmonic distortion feature vector contains information in multiple dimensions. The total harmonic distortion rate reflects the overall harmonic pollution level, and the odd harmonic proportion and even harmonic proportion further distinguish the source characteristics of harmonic components. The system uses the Z-score standardization method to process these three indicators to conform to the distribution with a mean of zero and a variance of one, facilitating subsequent model training and anomaly detection.

[0075] The phase shift feature vector directly uses the original value of the absolute value of the three-phase voltage phase difference without additional processing. This indicator directly reflects the synchronization state of the three-phase system and has important value for detecting phase imbalance or phase sequence errors.

[0076] The construction of the three-dimensional tensor is completed through channel splicing. The fluctuation rate feature, harmonic distortion feature, and phase shift feature occupy different channels of the tensor, forming a structured data representation. Subsequently, the system uses a 1x1 convolution kernel to reduce the dimension of the tensor, compressing the channel number from 3 to 32, generating the final multi-dimensional feature matrix. This matrix serves as the input of the subsequent anomaly detection model, carrying multi-dimensional information such as the dynamic characteristics of voltage fluctuation, harmonic pollution level, and phase synchronization state.

[0077] The entire dynamic feature extraction process runs in real-time on the edge computing node, with a calculation delay controlled at the millisecond level, meeting the needs of real-time grid monitoring. Through continuous processing of the sliding window, the system can continuously track the dynamic changes of the voltage waveform, providing data support for early detection and accurate positioning of abnormal events. The structured representation of the multi-dimensional feature matrix not only reduces the data dimension but also retains the key features of the original waveform.

[0078] In the implementation process, the system also considers the optimization of computing resources. For example, the FFT transformation uses a fast algorithm to reduce the amount of computation, and the phase-locked loop technology improves the phase detection accuracy through hardware acceleration. In addition, the step size and length of the sliding window are repeatedly verified to ensure a balance between computational efficiency and feature integrity. The design of the entire dynamic feature extraction module takes into account real-time, accuracy, and scalability, and can adapt to the monitoring needs of different scale power grids.

[0079] Embodiment 2: see Figure 3, the construction of the anomaly detection model and the implementation process of spatial correlation modeling involve historical data analysis, feature space modeling, real-time anomaly detection, and power grid topology correlation. The process is divided into two main stages: offline training and online detection. By combining the feature distribution of historical samples with real-time data stream analysis, accurate identification and spatial positioning of voltage sag, overvoltage, and harmonic resonance events are achieved.

[0080] In the offline training phase, the system extracts labeled voltage anomaly event samples from the historical database. The sag event sample includes event records where the voltage effective value suddenly drops below 90% of the rated value and lasts for 10 milliseconds to 1 minute; the overvoltage sample covers cases where the voltage rises above 110% of the rated value; and the harmonic resonance sample mainly records waveform data with significant amplification of specific harmonic components. After preprocessing, these samples form a data set containing complete waveform features before and after the event, with a time span of usually 5 minutes before the event to 5 minutes after the event.

[0081] The establishment of feature distribution boundaries uses multiple algorithms to adapt to the characteristics of different types of events. For sag events, the system uses a convex hull algorithm to construct the smallest enclosing region of sample points in a 32-dimensional feature space. This algorithm can automatically adapt to the shape of the feature point distribution, retaining a compact boundary that contains most normal samples. The boundary modeling of overvoltage events uses a probability density estimation method, which sets a reasonable probability threshold as the anomaly criterion by fitting the distribution of sample points in the feature space. Due to the complex characteristics of harmonic resonance events, the system uses a hierarchical clustering method to divide the samples into several sub-classes, and establishes a feature boundary for each sub-class.

[0082] In the online detection phase, the multi-dimensional feature matrix generated in real time is input into the anomaly detection module at 20 millisecond intervals. The system calculates the distance measure of each feature vector to the boundary of each type of event. For sag events, the Euclidean distance is used, for overvoltage events, the Mahalanobis distance is used to consider the correlation between features, and for harmonic resonance events, a specially designed spectral similarity index is used. When a feature vector breaks through the boundary of a certain type of event, the system records the specific coordinates and the breakthrough degree of the breakthrough point.

[0083] The activation of event feature recognition flags follows the dynamic threshold principle. The system counts the proportion of feature vectors that break through the boundary in the current time window. When the breakthrough proportion of sag events exceeds the pre-set threshold, the sag event recognition flag is triggered. At the same time, the system records key parameters such as the initial time, duration, and maximum breakthrough depth of the event. The determination of overvoltage events additionally considers the spatial aggregation of breakthrough points, requiring multiple adjacent feature vectors to continuously break through the boundary to confirm the occurrence of the event. The identification of harmonic resonance events focuses on the feature change pattern of specific harmonic frequency bands, and uses a pattern matching algorithm to improve detection accuracy.

[0084] The implementation of spatial correlation modeling relies on the deep integration of power grid topology data. The system obtains complete network connection relationships from the power grid GIS platform, including line impedance parameters, transformer ratio, circuit breaker status, and other information. These data are converted into a weighted graph structure, with nodes representing monitoring points and edges representing electrical connection relationships. The edge weights are calculated based on line impedance and electrical distance.

[0085] The calculation of time delay correlation coefficients uses a sliding window cross-correlation algorithm. The activation time series of event feature recognition markers are used as input, and the system searches for the time delay corresponding to the maximum correlation coefficient between adjacent nodes. The calculation process takes into account the physical characteristics of power grid signal propagation and sets a reasonable search range to avoid false correlations. Electrical connection relationship parameters are converted into normalized coupling coefficients based on line impedance, reflecting the electrical tightness between nodes.

[0086] The generation of spatial correlation weights combines information from two dimensions: time delay correlation and electrical connection degree. The system uses weighted summation to combine the two types of parameters, with time delay correlation coefficients reflecting the time characteristics of event propagation and electrical connection degrees embodying the physical structure characteristics of the network. The fused weight values are standardized to map to a unified range of 0 to 1.

[0087] The construction of directed connection edges follows the physical laws of power flow in the power grid. The system determines the direction of the edges based on the direction of power flow, and only connections with weights greater than a certain threshold are retained. The generation of the initial correlation graph uses an incremental update strategy, with the system automatically updating the connection relationships of relevant nodes whenever a new event is detected, maintaining the real-time nature of the graph. The edges in the graph not only contain weight information but also mark additional attributes such as event type, propagation time delay, and intensity attenuation, forming a complete power grid abnormal event propagation network.

[0088] The collaborative work of the anomaly detection model and spatial correlation modeling achieves a leap from single-point anomaly detection to full-network event analysis. The system can distinguish between local device faults and global power grid disturbances, identify the origin location and propagation path of abnormal events. Through continuous learning and adaptive updating mechanisms, the model can gradually improve feature boundaries and correlation rules, adapting to changes in power grid structure and the emergence of new abnormal patterns. The entire implementation process focuses on balancing computational efficiency and actual engineering needs, ensuring reliable monitoring performance in complex power grid environments.

[0089] In terms of implementation details, the system employs various optimization measures to improve operational efficiency. The storage of historical samples uses a time series database, supporting fast retrieval and batch processing. The feature distance calculation uses vectorized operations to accelerate, and the GPU parallel processing capability is used to shorten the response time. The update of spatial correlation weights uses an incremental calculation strategy, and only the affected local area is recalculated. These designs enable the system to meet the stringent time delay requirements of large-scale power grid real-time monitoring, providing effective technical support for safe operation of the power grid.

[0090] The adaptive ability of the model is realized through a regular update mechanism. The system sets automatic trigger conditions, and when significant drift in feature distribution or new abnormal patterns are detected, the model parameter adjustment process is started. The update process uses online learning algorithms to gradually adjust the feature boundaries and correlation rules without interrupting real-time monitoring. Historical model versions are also preserved for result comparison and fault backtracking when necessary. This design ensures the timeliness of the model while maintaining the stability of the system.

[0091] The processing of power grid topology data takes into account various special situations in actual engineering. For complex connection relationships such as double-circuit lines and parallel transformers, the system establishes equivalent models to accurately reflect electrical characteristics. Different voltage level areas within substations are modeled as a hierarchical structure, and virtual nodes are used to implement cross-voltage level correlation analysis. These refined modeling methods improve the accuracy of spatial correlation analysis, enabling the system to identify hidden relationships that traditional methods cannot detect.

[0092] The spatio-temporal correlation analysis of abnormal events not only considers electrical connection relationships, but also introduces spatial information such as line direction and device location provided by geographic information systems. The system can visually display the propagation process of abnormal events, assisting operators in intuitively understanding the development trend of power grid disturbances. This multi-dimensional correlation analysis enhances the comprehensiveness of abnormal diagnosis, providing more comprehensive information support for subsequent fault handling decisions.

[0093] The entire implementation process focuses on seamless integration with existing power grid automation systems. Abnormal detection results are uploaded to the energy management system through standard protocols, and spatial correlation maps are shared with the SCADA system display interface. The system supports multiple data export formats, facilitating data interaction with other analysis tools.

[0094] Example 3: see Figure 4The generation of the multi-parameter association graph and the workflow of the diagnosis model are based on the initial association graph. Through the fusion of power grid load distribution, environmental parameters, and graph neural network technology, the system realizes the deep analysis and accurate diagnosis of abnormal events. The initial association graph contains directed connection edges generated by spatial correlation modeling, and each edge has a weight coefficient reflecting the intensity of event propagation. On this basis, the system performs multi-dimensional data fusion and graph optimization, and finally outputs a multi-parameter association graph with weight annotations, providing structured input for power grid state diagnosis.

[0095] The introduction of load distribution parameters expands the analysis dimension of the association graph. The system obtains the active power, reactive power, and power factor data of each node from the energy management system in real time, with a sampling interval of 15 minutes. These parameters are normalized and converted into load rate indicators ranging from 0 to 1, representing the operating state of the node. Environmental temperature and humidity data are collected by IoT sensors deployed at substations and line terminals. The temperature measurement range covers -40°C to +85°C, and the humidity measurement range covers 0% to 100% RH. Environmental parameters are updated every 5 minutes, and after sliding average filtering, transient fluctuation interference is eliminated.

[0096] The architecture of the graph neural network is designed using the GraphSAGE framework, which is suitable for processing dynamic changes in power grid topology. The node feature vector contains three parts: load features, environmental features, and event features. Load features consist of the current load rate and its first-order difference, reflecting the load change trend; environmental features include temperature, humidity, and their gradient changes; event features record the types and intensity of detected abnormalities. The aggregation layer of the network uses the mean pooling function, with 3 iterations to ensure sufficient propagation of node information without excessive smoothing. In each iteration, nodes aggregate feature information from their neighbors and update their own feature representation. After multiple layers of propagation, each node obtains a 64-dimensional coupled feature vector containing local topological information.

[0097] The attention weighting mechanism is used to fuse the spatial correlation weight and the coupled feature vector. The system uses a multi-head attention structure, setting 4 independent attention heads to calculate in parallel. The query vector is taken from the spatial correlation weight matrix, and the key and value vectors are obtained by linear transformation of the coupled features. The calculation of attention weights follows the following formula:

[0098]

[0099] where, represents the attention weight of node i to node j, is the query vector of node i, is the key vector of node j, The dimension of the key vector. The formula calculates the relative importance between nodes, and the attention weight is fused with the original space correlation weight in a preset proportion to generate the updated edge weight coefficient. This mechanism enables the system to adaptively adjust the importance of the event propagation path and highlight the key connection relationship.

[0100] The probability reordering process uses an improved PageRank algorithm, which introduces the influence of edge weight based on the traditional algorithm. The algorithm sets the damping factor to 0.85 and iteratively calculates the global importance score of the node. During the sorting process, the system preferentially retains paths with high cumulative weight, which usually correspond to the main propagation channels of abnormal events. The extraction of key path node sequences uses a variant Dijkstra algorithm to search for the top 5 paths with the largest weight accumulation value in the graph, with a path length limit of 6 nodes to avoid excessive extension.

[0101] The input features of the diagnostic model are carefully extracted from the multi-parameter correlation graph. The voltage recovery time parameter records the duration of the event feature recognition mark from activation to disappearance, reflecting the self-healing ability of the power grid. The harmonic attenuation rate is calculated by fitting the slope of the total harmonic distortion curve within 30 seconds after the event ends, representing the response characteristics of the filtering device. The phase synchronization deviation calculates the standard deviation of the three-phase phase difference of all monitoring points on the key path, indicating the system synchronization stability. These parameters are processed by difference and moving average smoothing to form a 15-dimensional feature vector input into the decision tree classifier.

[0102] The design of the decision tree classifier takes into account the characteristics of the power grid diagnosis task. In addition to the original parameters, the first-order difference and moving average values are also included to enhance the expression of time series features. The classifier uses Gini impurity as the splitting criterion, with a maximum tree depth limit of 8 layers to prevent overfitting. The output layer generates a probability distribution of three diagnostic results: the insulation aging risk level is divided into low, medium, and high levels; the capacitor compensation failure probability is expressed as a percentage; and the transformer overload warning uses a binary flag. The training process of the classifier uses a class weight adjustment strategy to balance the frequency of different diagnostic results.

[0103] The dynamic updating mechanism of the multi-parameter correlation graph ensures the timeliness of the analysis. When a new event is detected or the environmental parameters change significantly, the system triggers a local recalculation process. Only the nodes and edges in the affected area are updated for features and weights, while the rest remain stable. This incremental updating strategy significantly reduces computational overhead, enabling the system to maintain a response speed of seconds. The graph version management uses a snapshot mechanism, saving the complete state every hour to support historical backtracking and comparative analysis.

[0104] The visualization subsystem converts the complex correlation graph into an intuitive graphical display. The node color coding represents the abnormal type, and the size reflects the load rate level; the edge width is proportional to the weight coefficient, and the arrow direction indicates the event propagation path. The three-dimensional view superimposes geographical information, clearly presenting the spatial distribution and propagation trend of abnormal events. The operating personnel can focus on the concerned area through the interactive interface and view the detailed parameters and event records of any node. The visualization system also provides an automatically generated analysis report, summarizing the current power grid state and potential risk points.

[0105] Integration with the power grid control system enables closed-loop management. The diagnosis results are transmitted to the protection devices and automatic control system through the IEC61850 protocol. When the transformer overload risk is detected, the system sends load adjustment suggestions to the relevant substation; for capacitor compensation failure warning, it triggers the state checking instruction of the reactive power compensation equipment. The generation of control instructions considers multi-objective optimization, minimizing the impact on power supply quality while eliminating abnormalities. All control operations are recorded with detailed logs, supporting post-analysis and process tracing.

[0106] The system is implemented using a microservices architecture, with each functional module deployed independently. The graph generation service runs on servers with GPU acceleration, handling computationally intensive tasks; the diagnosis model is deployed on edge computing nodes for low-latency response; the visualization front-end supports access from multiple terminal devices. The message bus uses a publish / subscribe model to ensure efficient data flow between modules. This architecture design meets real-time requirements while maintaining system scalability and fault tolerance.

[0107] Example 4: Refer to Figure 5 , the implementation process of dynamic sampling frequency adjustment and diagnosis model training is demonstrated through a specific case. In the monitoring system of a certain 220kV substation, the voltage sensor array detects a harmonic abnormality on the medium voltage side of the transformer. The diagnosis model generates a diagnosis report based on the analysis results of the multi-parameter correlation graph, triggering the parameter adaptive adjustment process.

[0108] When the diagnosis report contains a transformer overload warning identifier, the system initiates a hierarchical sampling frequency adjustment strategy. The associated sensor nodes include the transformer high, medium, and low voltage sides, totaling 6 monitoring points, with a baseline sampling frequency originally set at 1kHz. The adjustment process gradually increases the frequency by 1kHz every 5 minutes, up to a preset upper limit of 10kHz. During the adjustment process, the system monitors the storage space and network bandwidth usage in real-time to avoid affecting other functions due to a surge in data volume. The filter parameters associated with the transformer overload warning are adjusted simultaneously, with the bandpass filter center frequency initially set at the detected 5th harmonic frequency of 250Hz and the bandwidth set at ±2Hz. The filter coefficients are recalculated every 60 seconds, dynamically optimized based on the spectral analysis results of harmonic components in the last 1 minute.

[0109] The training data of the diagnostic model comes from the historical event records of a provincial power grid over the past three years, including various types of voltage abnormal events and their corresponding power grid device states. The training samples are stratified sampled according to the event type and severity to ensure the balance of data distribution. The following table shows some key features of the training samples:

[0110] Event type Duration (ms) Harmonic distortion rate (%) Phase deviation (degrees) Load rate (%) Temperature (°C) Diagnostic label Sag event 120 3.2 5.8 65 28 Capacitance compensation failure Overvoltage event 80 1.5 2.1 92 31 Transformer overload warning Harmonic resonance event 300 8.7 7.3 78 25 Insulation aging risk Composite event 200 5.4 9.2 85 30 Multiple fault warning

[0111] The architecture of the bidirectional long short-term memory network is optimized for the characteristics of power grid time series data. The input sequence length is set to 60 time steps, corresponding to a 60-minute time span of the multi-parameter correlation graph. The network hidden layer contains 128 neurons, using the tanh activation function to capture the time series dependence. The input features include event duration, harmonic distortion rate, phase deviation, and other dynamic parameters, as well as load rate, environmental temperature, and other state parameters. The network training uses a sliding window strategy with a window step size of 10 time steps to increase sample diversity.

[0112] The design of the multilayer perceptron focuses on efficient fusion of features. The number of neurons in the three fully connected layers is 64, 32, and 16, respectively, with ReLU activation functions and Dropout regularization between layers. The output layer corresponds to three diagnostic tasks: the insulation aging risk level uses a three-class Softmax output; the capacitor compensation failure probability uses a Sigmoid function to output a value between 0 and 1; and the transformer overload warning uses a binary output. The loss function is designed as the weighted sum of the three tasks, with a higher weight coefficient for the overload warning task to reflect its priority in power grid safety.

[0113] The model training process uses an adaptive optimization algorithm, with an initial learning rate of 0.001 and a batch size of 256. During training, the loss on the validation set is monitored, and when there is no significant decrease for 10 consecutive epochs, the early stopping mechanism is triggered. The model saving uses the optimal checkpoint strategy, retaining the parameter version with the best validation set performance. After training, the model is compressed in volume through quantization technology to facilitate deployment on edge computing devices.

[0114] The implementation of the parameter adaptive adjustment module considers the actual engineering constraints. The increase in sampling frequency needs to consider the limit performance of the sensor hardware, and after reaching the maximum sampling rate, the filter order is increased. The band-stop filter is implemented using a cascade of second-order sections, and only the filter sections corresponding to the affected frequency bands are updated in each adjustment period. The adjustment instructions are sent to the sensor terminal through an encrypted channel, and after the terminal device returns an acknowledgement signal, the system records the parameter change log.

[0115] A continuous learning mechanism is established after the deployment of the diagnosis model. New event cases are extracted from the power grid operation database every month, and after data cleaning and labeling, they are added to the training set. The model uses an online learning algorithm for incremental training, and after each update, an A / B test is performed with the original model to compare the diagnosis accuracy on the latest data. The system retains the last three versions of the model, and when the performance of the new version decreases, it can quickly roll back.

[0116] The fault case backtracking function helps to improve the parameter adjustment strategy. The system records the power grid state data before and after each parameter adjustment, and when a related fault occurs later, it can analyze the adjustment effect. A 110kV line fault analysis shows that after the harmonic anomaly warning, increasing the sampling frequency to 8kHz helps to detect insulation deterioration signs 10 minutes earlier. Such cases are converted into optimization of adjustment rules after expert evaluation, for example, shortening the interval time of sampling frequency increase in specific working conditions.

[0117] The integration with existing power grid management systems uses a loose coupling method. Diagnosis reports are uploaded to the production management system through standardized interfaces, and parameter adjustment instructions are transmitted to field devices through a security gateway. The system defines clear interface protocols, including data format, communication timing, and exception handling processes. During the implementation process in the pilot substation, this scheme successfully achieved data interoperation with traditional SCADA systems and device management systems.

[0118] The hardware resource configuration scheme dynamically adjusts according to the monitoring scale. For hub substations, a dedicated edge server is deployed to run the diagnosis model; for ordinary substations, containerization technology is used to share regional computing resources. The firmware of sensor terminals supports remote upgrading, and when a new type of filtering algorithm is developed, it can be updated in batches. The storage system uses a hierarchical design, with high-frequency sampling data cached locally for 6 hours before being archived at a lower frequency, and key feature data saved permanently.

[0119] This implementation has shown good adaptability during the trial operation in a certain regional power grid. In the face of voltage fluctuations during the summer load peak period, the system accurately identifies capacitor bank failure cases and detects multiple potential faults in advance by adjusting monitoring parameters. Under low temperature conditions in winter, the environmental temperature compensation algorithm effectively eliminates the interference of temperature measurement errors on the diagnosis results. Operation log analysis shows that the parameter self-adaptive adjustment strategy reduces the average invalid data storage volume by 30% while ensuring monitoring accuracy.

[0120] The system maintenance establishes a standardized process. Daily automatic checks of parameter synchronization status for each sensor node, weekly generation of device health reports, and monthly performance evaluation of the model are performed. The maintenance interface provides intuitive operation guidelines, supporting remote reset of abnormal nodes, manual override of adjustment parameters, and other emergency operations. The knowledge base includes typical fault handling cases to assist operations personnel in quickly solving common problems.

[0121] User feedback mechanism continuously optimizes system behavior. Operators can score the diagnostic reports and parameter adjustment effects, and the system collects these feedbacks for improving decision logic. A user's feedback that "overload warning is too sensitive" prompts the development team to recalibrate the weight coefficient of transformer temperature parameter. This closed-loop improvement process gradually adapts the system to the operation characteristics and operation habits of the local power grid.

[0122] Embodiment 5: Triggering and execution process of feature distribution boundary dynamic expansion operation The mechanism establishes a supplementary path outside the regular anomaly detection process. When real-time monitoring data shows a feature pattern beyond historical cognition, the system automatically starts the boundary expansion program to avoid missed detection due to model limitations.

[0123] The triggering of dynamic expansion operation is based on continuous feature space monitoring. The system tracks the relative position of each vector point in the multi-dimensional feature matrix and the existing boundary in real time, calculates the proportion and distribution density of boundary-breaking points in each detection period. When the breakthrough proportion of a certain type of feature vector exceeds the preset threshold for three consecutive periods, and these breakthrough points show an aggregated pattern in the feature space rather than random scattering, the system determines that a new type of abnormal pattern may occur. The triggering condition considers the spatial distribution characteristics of breakthrough points, requiring at least one breakthrough area containing several adjacent feature vectors to exclude the influence of isolated noise points.

[0124] The extraction of abnormal feature vector set uses density clustering method. The system performs neighborhood analysis on all boundary-breaking points in the current detection period to identify areas with dense data points in the feature space. The clustering process sets a reasonable neighborhood radius parameter to ensure that breakthrough points with similar features are classified into the same category, while filtering out sparse abnormal values. Each cluster represents a potential new type of abnormal pattern, and the system records key parameters such as the center position, distribution range, and the number of feature vectors contained in each cluster.

[0125] The expansion of feature distribution boundary is based on the convex hull algorithm. The system calculates the convex hull boundary of all feature vectors contained in each identified cluster. This boundary is expanded outward in the original feature space dimensions by a preset proportion, and the expansion proportion varies according to the physical meaning of different feature dimensions. The expansion of voltage fluctuation-related dimensions is relatively conservative to avoid misjudging normal fluctuations as abnormal; the expansion of harmonic feature dimensions is larger to accommodate possible atypical harmonic combinations; the expansion of phase feature dimensions considers the requirements of power grid synchronization stability and sets asymmetric expansion intervals.

[0126] The boundary updating process adopts an incremental learning strategy. The system retains the original feature distribution boundary of the event type and only adds an extended region for newly identified abnormal patterns. To prevent excessive expansion of the boundary from causing a rise in false detection rates, the range of each expansion is limited to the minimum necessary. The updated boundary parameters are managed through a version control mechanism, and each change is recorded with detailed expansion reasons, involved feature dimensions, and adjustment amplitudes. The system maintains a history of boundary changes and supports reverting to previous versions if necessary.

[0127] The synchronized updating of model parameters adopts a distributed transaction mechanism. After receiving new feature distribution boundary parameters, the edge computing node first verifies its compatibility with the existing model in an isolated environment. After verification, the new parameters are updated to the running anomaly detection model in an atomic operation, ensuring that all nodes switch to the new version at the same time. The synchronization process takes into account network delays and device heterogeneity, setting a reasonable timeout and retry mechanism to ensure eventual consistency.

[0128] Analysis reports for new events are automatically generated. The system performs categorical analysis on the feature patterns involved in each boundary expansion operation, extracting key features that distinguish them from known abnormal types. The report includes visual projections of the feature space, similarity comparisons with various known events, and possible physical phenomenon analysis. These reports are presented through standardized templates for reference by power grid analysts and as a basis for subsequent model optimization.

[0129] The dynamic expansion mechanism is self-monitored to ensure system stability. The system continuously tracks changes in model performance after boundary expansion, including detection sensitivity, false alarm rate, and other key indicators. When the expanded boundary is not effectively utilized for a long time in subsequent monitoring or causes a significant increase in the false alarm rate of a certain type of event, the boundary contraction mechanism is triggered. This self-regulating ability keeps the system within the optimal detection accuracy range, avoiding performance degradation due to excessive expansion.

[0130] The handling process of unknown type events forms a closed-loop management. When a new abnormal pattern is identified and the boundary is expanded, the system marks it as a pending state, and the relevant monitoring data automatically enters the focus queue. Grid operation personnel can manually review these cases to determine whether they are real faults or new disturbances. Verified cases are added to the historical sample library for formal training of subsequent models, while false patterns trigger boundary adjustment to optimize detection rules.

[0131] The system takes into account the complexity of the actual engineering environment. The speed and amplitude of the feature boundary expansion are adjustable parameters that adapt to the monitoring needs of power grids of different voltage levels. High-voltage power grids use a relatively conservative expansion strategy, focusing on stability; low-voltage distribution networks allow more flexible boundary adjustments to improve response speed to new phenomena. Edge nodes with limited device resources support a simplified expansion algorithm that reduces computational overhead while ensuring core functionality.

[0132] The linkage with the power grid knowledge base enhances the rationality of boundary expansion. Before deciding to expand the feature boundary, the system queries structured knowledge such as power grid device parameters and historical operation records to assess the relevance of new feature patterns to known power grid phenomena. This domain knowledge-based verification step effectively distinguishes between real abnormal new patterns and interference factors such as measurement noise, improving the credibility of expansion decisions.

[0133] Long-term operation and maintenance establish standardized procedures. The system regularly generates feature boundary usage reports, which statistically analyze the actual detection effectiveness of various boundaries. Operation and maintenance personnel can view boundary change history to understand the system's adaptive learning process. The diagnostic interface provides hierarchical warnings for new abnormal patterns, using different prompt methods based on confidence levels to ensure that no abnormalities are missed and to avoid excessive interference with operators.

[0134] The implementation of dynamic expansion mechanisms enables the system to have continuous evolution capabilities. As the power grid structure changes, new equipment is put into operation, and the characteristics of electricity load change, the monitoring system can automatically adjust the detection standards to keep pace with the real power grid environment. This adaptive feature significantly extends the effective service period of the system and reduces the need for upgrading due to outdated detection models. In actual operation, the system triggers one to two boundary expansion operations every quarter, and each adjustment effectively covers new abnormal patterns, maintaining stable monitoring performance levels.

[0135] User interaction design focuses on transparency and controllability. Operation personnel can view the detailed parameters of all current feature distribution boundaries at any time to understand the system's judgment criteria for various abnormal events. The trigger records and decision basis of boundary expansion operations are saved completely to support post-review. The system provides a manual intervention interface that allows experienced engineers to make necessary adjustments based on automatic decisions, achieving intelligent monitoring through human-machine collaboration.

[0136] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0137] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for real-time monitoring of multiple parameters of a voltage sensor in an Internet of Things environment, characterized in that: The multi-channel voltage sensor array deployed at the grid node collects real-time voltage fluctuation data in the target area and transmits the collected raw voltage data stream to the edge computing node; Perform dynamic feature extraction on the original voltage data stream within the edge computing node to generate a multi-dimensional feature matrix containing voltage fluctuation rate features, harmonic distortion features, and phase offset features; Analyzing the changing trends of various eigenvectors in the multidimensional feature matrix based on a preset anomaly detection model to identify voltage sag event characteristics, overvoltage event characteristics, and harmonic resonance event characteristics; Based on the grid topology parameters, the spatial correlation modeling of the identified voltage sag event characteristics, overvoltage event characteristics, and harmonic resonance event characteristics is performed. The parameter coupling strength between adjacent sensor nodes is calculated to generate a multi-parameter correlation map that includes the event propagation path and impact range. The multi-parameter correlation map is input into a diagnostic model to generate a power grid status diagnostic report, and the sampling frequency and filtering parameters of the voltage sensor are dynamically adjusted according to the content of the diagnostic report.

2. The method for real-time monitoring of multiple parameters of voltage sensors in an Internet of Things environment according to claim 1 is characterized in that: The specific process of the dynamic feature extraction operation includes: The original voltage data stream is segmented using a sliding time window. The voltage RMS gradient, the energy ratio between the fundamental component and the harmonic component, and the absolute value of the three-phase voltage phase difference are simultaneously calculated within each time window. Compare the voltage RMS change gradient with a preset reference fluctuation threshold to generate a normalized fluctuation rate feature vector; Constructing harmonic distortion feature vectors based on harmonic component energy ratios and absolute values ​​of phase differences; The normalized volatility eigenvector, harmonic distortion eigenvector and phase shift eigenvector are concatenated into a three-dimensional tensor structure, and the convolution kernel is used to compress the channel dimension to generate a multidimensional feature matrix.

3. The method for real-time monitoring of multiple parameters of voltage sensors in an Internet of Things environment according to claim 2, characterized in that: The anomaly detection model is constructed as follows: Extract sag event samples, overvoltage event samples, and harmonic resonance event samples marked in historical voltage data; Calculating the characteristic distribution intervals of various event samples on the multidimensional characteristic matrix respectively, and establishing characteristic distribution boundaries of sag events, overvoltage events, and harmonic resonance events; When vector points exceeding a preset ratio in the multi-dimensional feature matrix generated in real time break through the feature distribution boundary, an event feature identification flag of a corresponding type is activated.

4. The method for real-time monitoring of multiple parameters of a voltage sensor in an Internet of Things environment according to claim 3 is characterized in that: The operation steps of the spatial correlation modeling include: Obtaining electrical connection relationship parameters of each sensor node in the power grid topology; Calculate the time delay correlation coefficient of event feature identification marks between adjacent nodes, and weight the time delay correlation coefficient with the electrical connection relationship parameter to generate the spatial correlation weight; Directed connection edges between nodes are constructed based on spatial association weights to form an initial association graph that includes the propagation direction and weight of event features.

5. The method for real-time monitoring of multiple parameters of a voltage sensor in an Internet of Things environment according to claim 4 is characterized in that: The multi-parameter association map is generated as follows: Overlaying grid load distribution parameters and ambient temperature and humidity parameters on the initial correlation map; A graph neural network is used to aggregate node load distribution parameters and environmental temperature and humidity parameters to generate a load-environment coupling feature vector. The load-environment coupling feature vector and the spatial association weight are weighted, and the weight coefficient of the directed connection edge is updated; The event feature propagation paths are probabilistically reordered according to the updated weight coefficients to generate a multi-parameter association map with weight annotations.

6. The method for real-time monitoring of multiple parameters of a voltage sensor in an Internet of Things environment according to claim 5, characterized in that: The workflow of the diagnostic model includes: Analyze the topological structure of event feature propagation paths in multi-parameter correlation graphs and extract key path node sequences; Calculate the voltage recovery time parameters, harmonic attenuation rate parameters and phase synchronization deviation parameters of the critical path node sequence; The above parameters are input into the pre-trained decision tree classifier to output the grid insulation aging risk level, capacitor compensation failure probability and transformer overload warning sign.

7. The method for real-time monitoring of multiple parameters of voltage sensors in an Internet of Things environment according to claim 1, characterized in that: The method of dynamically adjusting the sampling frequency and filtering parameters of the voltage sensor is as follows: When the diagnostic report contains a transformer overload warning indicator, the sampling frequency of the associated sensor node is increased to the preset maximum threshold; When the diagnostic report shows that the harmonic attenuation rate is lower than the reference value, activate the adaptive band-stop filter and configure its center frequency parameters.

8. The method for real-time monitoring of multiple parameters of a voltage sensor in an Internet of Things environment according to claim 6, characterized in that: The training process of the diagnostic model includes: Collect historical multi-parameter correlation maps and their corresponding grid fault diagnosis labels; Constructing a bidirectional long short-term memory network to encode the topological features of multi-parameter association maps; Generate a predictive diagnosis report by decoding topological features through a multi-layer perceptron; Calculate the cross entropy loss value between the prediction diagnosis report and the power grid fault diagnosis label, and backpropagate to update the network parameters.

9. The method for real-time monitoring of multiple parameters of a voltage sensor in an Internet of Things environment according to claim 3, characterized in that: Also includes: When the type of the activated event feature identification flag in the real-time generated multi-dimensional feature matrix exceeds the range of historical samples, the feature distribution boundary dynamic expansion operation is performed: Extract the set of abnormal feature vectors that break through the feature distribution boundary; Taking the convex hull boundary of the abnormal feature vector set as a reference, the characteristic distribution boundary of the sag event, the characteristic distribution boundary of the overvoltage event and the characteristic distribution boundary of the harmonic resonance event are expanded according to a preset ratio; Update the feature distribution boundary parameters of the anomaly detection model and synchronize them to the edge computing nodes.

10. A device for real-time monitoring of multiple parameters of a voltage sensor in an Internet of Things environment, used to implement the method for real-time monitoring of multiple parameters of a voltage sensor in an Internet of Things environment according to any one of claims 1 to 9, characterized in that: include: A multi-channel voltage data acquisition unit, which is used to collect raw voltage data streams through a sensor array and transmit them to edge computing nodes; A dynamic feature extraction unit is used to generate a multi-dimensional feature matrix containing voltage fluctuation rate features, harmonic distortion features, and phase offset features within the edge computing node; An abnormal event recognition unit, used to identify voltage sag event characteristics, overvoltage event characteristics, and harmonic resonance event characteristics based on an abnormality detection model; A spatial correlation modeling unit is used to generate a multi-parameter correlation map including event propagation paths and impact ranges based on the grid topology parameters; A power grid diagnosis unit, used to output a power grid status diagnosis report based on a multi-parameter correlation map; The parameter adaptive adjustment unit is used to dynamically control the sampling frequency and filtering parameters of the voltage sensor according to the diagnosis report.

Images