A computer system based on power internet cloud edge computing power collaborative computing model

Abstract

The application discloses a kind of based on electric power internet of things cloud edge computing power collaborative computing model computer system, it is suitable for electric energy quality disturbance event monitoring, the computer system includes edge side time period determination module, edge side frequency domain analysis module, cloud event classification module, cloud source node determination module and edge side monitoring module, wherein the edge side time period determination module is used to determine disturbance event time period by machine learning model;The edge side frequency domain analysis module is used to obtain abnormal analysis data by calculation;The cloud event classification module is used to determine disturbance event type by K-means clustering algorithm;The cloud source node determination module is used to determine actual source node by preset graph neural network;The edge side monitoring module is used to establish power grid disturbance event early warning monitoring point position layout to the actual source node.This application can improve the real-time and accuracy of electric energy quality disturbance event monitoring.

Description

Technical Field

[0001] This invention relates to the field of smart grid technology, and in particular to a computer system based on a cloud-edge computing collaborative computing model of the power Internet of Things. Background Technology

[0002] In the context of the power Internet of Things (IoT), rapid load switching, grid connection of new energy sources, and the extensive integration of various nonlinear devices cause power quality indicators such as grid voltage, frequency, and harmonics to be constantly changing dynamically. If disturbances (such as voltage drops, frequency shifts, and harmonic exceedances) are not detected and traced in a timely manner, they can lead to anything from minor malfunctions of sensitive equipment and production line shutdowns to cascading failures, expanding the scope of power outages, and causing significant economic losses and social impact. Therefore, it is essential to monitor power quality disturbances in real time and accurately to pinpoint the source of the disturbance, assess its type and evolution trend in the shortest possible time, and optimize early warning deployment and operational control strategies accordingly to ensure the safe, high-quality, and economical operation of the power grid.

[0003] Existing technologies typically employ edge device monitoring or cloud-based deep analysis. However, edge devices have limited computing power, enabling them to perform only simple threshold exceedance detection or short-time Fourier transforms, making it difficult to conduct high-precision, wide-bandwidth frequency domain deep analysis locally. While the cloud possesses powerful computing capabilities, it is limited by transmission latency and bandwidth, hindering real-time responses to transient disturbances and leading to missed detections or delays. Furthermore, traditional methods use fixed-length analysis windows, which cannot be dynamically adjusted according to the duration of the disturbance, easily resulting in waveform truncation and loss of features. In utilizing historical data, they simply compare thresholds or waveform similarity, lacking the ability to model the evolution of disturbance types and regional migration, making it difficult to quickly relocate the source of the disturbance when its location changes.

[0004] Therefore, there is an urgent need to invent a cloud-edge computing collaborative computing model that integrates the instantaneous characteristics and long-term evolution trends of disturbance events to achieve rapid location and accurate identification of disturbance sources. Summary of the Invention

[0005] This invention provides a computer system based on a cloud-edge computing collaborative computing model of the power Internet of Things, which can improve the real-time performance and accuracy of power quality disturbance event monitoring through cloud-edge computing collaboration.

[0006] In a first aspect, embodiments of the present invention provide a computer system based on a cloud-edge computing collaborative computing model of the power Internet of Things, suitable for monitoring power quality disturbance events. The computer system integrates an edge-side time period determination module, an edge-side frequency domain analysis module, a cloud-side event classification module, a cloud-side source node determination module, and an edge-side monitoring module.

[0007] The edge-side time period determination module is used to calculate the real-time disturbance intensity level based on the first voltage signal sequence and the first frequency signal sequence of the target power grid collected in real time, through a preset machine learning model. When the real-time disturbance intensity level exceeds a preset intensity threshold, the disturbance event time period is determined based on the real-time disturbance intensity level.

[0008] The edge-side frequency domain analysis module is used to adjust the width of the analysis window according to the time period of the disturbance event, so as to extract the second voltage signal sequence and the second frequency signal sequence through the analysis window, and to perform frequency domain feature extraction based on the second voltage signal sequence and the second frequency signal sequence to obtain anomaly analysis data.

[0009] The cloud-based event classification module is used to calculate the similarity between the anomaly analysis data and each historical disturbance record in the pre-stored historical disturbance record set to obtain a first similarity. Based on the first similarity, the historical disturbance record set is filtered to obtain a subset of historical disturbance records. Then, the disturbance event type is determined based on the subset of historical disturbance records and the anomaly analysis data using a preset K-means clustering algorithm.

[0010] The cloud-based source node determination module is used to obtain a list of candidate nodes for disturbance sources through a preset graph neural network based on pre-stored target power grid topology data and the disturbance event type, and to calculate the second similarity between each candidate node and the pre-stored historical source nodes, so as to determine the actual source node based on the second similarity; wherein, the historical source nodes and each candidate node are of the same disturbance event type;

[0011] The edge-side monitoring module is used to perform real-time frequency monitoring on the actual source node to obtain a real-time frequency value sequence, generate disturbance event identification tags based on the real-time frequency value sequence, and establish a power grid disturbance event early warning monitoring point layout based on the disturbance event identification tags.

[0012] This invention utilizes an edge-side time period determination module to collect voltage and frequency signals in real time. A machine learning model calculates the disturbance intensity level, first quantifying the disturbance intensity to filter out minor fluctuations and avoid invalid data processing at the edge, thus reducing computational power consumption. Leveraging the high transmission speed of edge devices, the time window for disturbance events is quickly identified, defining an effective analysis window for subsequent frequency domain analysis. This prevents feature distortion caused by a large amount of normal data being mixed into the frequency domain analysis, serving as the time reference for all subsequent analyses. The edge-side frequency domain analysis module dynamically adjusts the window width to ensure complete coverage of the disturbance waveform, solving the problems of "missed disturbance sampling" or "excessive noise sampling" in fixed windows. Frequency domain extraction converts the time-domain signal into quantifiable frequency-domain features, combining them with anomaly markers to generate "anomaly analysis data," providing a feature carrier for cloud-based classification and realizing the core value of real-time preliminary analysis at the edge. The cloud-based event classification module utilizes cloud computing power to process massive amounts of historical data. Through similarity filtering and K-means clustering, it addresses the issue of insufficient edge computing power hindering in-depth analysis. Precisely classifying disturbance types provides a basis for subsequent source location. The cloud-based source node determination module, combining grid topology and disturbance type, uses graph neural networks for reverse tracing to quickly identify candidate nodes, avoiding blind grid-wide searches. A second similarity method accurately selects the actual source from the candidates, solving the problem of difficulty in locating disturbances due to migration, thereby improving the source location rate. The edge-side monitoring module continuously monitors the located sources, generating identification tags to support full disturbance information recording. Early warning points covering the source and propagation path are established according to the tags, forming a closed-loop optimization. Subsequent similar disturbances can be quickly detected through these early warning points, improving response speed and realizing the value of one-time location and long-term early warning. Compared with existing technologies, this invention can improve the real-time performance and accuracy of power quality disturbance event monitoring through cloud-edge-device computing power collaboration.

[0013] Furthermore, the step of calculating the real-time disturbance intensity level based on the first voltage signal sequence and the first frequency signal sequence of the target power grid acquired in real time, using a preset machine learning model, specifically involves:

[0014] The first voltage signal sequence and the first frequency signal sequence of several sampling points in the target power grid are acquired in real time through a preset sliding time window.

[0015] Based on the first voltage signal sequence, the difference between the real-time effective voltage value and the rated voltage value at each sampling point is calculated sequentially to obtain the voltage amplitude deviation sequence;

[0016] Based on the first frequency signal sequence, the difference between the real-time frequency value and the standard frequency value at each sampling point is calculated sequentially to obtain the frequency offset sequence;

[0017] Based on the voltage amplitude deviation sequence, frequency offset sequence, preset deviation anomaly threshold, and preset offset anomaly threshold, the number of abnormal sampling points is counted to obtain anomaly statistics. Then, based on the anomaly statistics, the real-time disturbance intensity level is calculated using a preset neural network model.

[0018] This invention employs a sliding time window to acquire signals, achieving "continuous real-time sampling" of voltage and frequency signals. This ensures that no instantaneous disturbances are missed, providing a continuous data foundation for subsequent deviation calculations. By calculating voltage amplitude deviation sequences and frequency offset sequences, the time-domain signals are transformed into quantified deviation indicators, avoiding qualitative ambiguity in the judgment of disturbances and providing a quantitative basis for anomaly statistics. By statistically analyzing abnormal sampling points, the deviation sequence is transformed into anomaly degree data, avoiding misjudgment based on a single sampling point. Furthermore, neural networks excel at processing multi-dimensional data, resulting in more accurate intensity level calculations and reducing the false triggering of subsequent actions.

[0019] Furthermore, determining the time period of the disturbance event based on the real-time disturbance intensity level specifically involves:

[0020] Perform a differential operation on the voltage amplitude deviation sequence of each sampling point to obtain the deviation change rate, and take the sampling point where the deviation change rate exceeds the preset change rate threshold as the abrupt change point;

[0021] Based on the voltage amplitude deviation sequence corresponding to the mutation point, the disturbance start point and disturbance end point are identified, and the maximum value of the voltage amplitude deviation sequence between the disturbance start point and the disturbance end point and the peak value of the frequency offset sequence are extracted.

[0022] Based on the real-time disturbance intensity level and the preset weighting coefficient, the maximum value and the peak value are weighted and summed to obtain a comprehensive disturbance index. If the comprehensive disturbance index is greater than the preset comprehensive disturbance threshold, the disturbance event time period is determined based on the disturbance start point and the disturbance end point.

[0023] This invention accurately locates the start and end points of disturbances by calculating the rate of change of deviation, avoiding the inclusion of too much normal data in the time period due to inaccurate judgment of the start and end points, thus improving accuracy while reducing computational load. By extracting the maximum voltage value and the peak frequency value, a weighted comprehensive disturbance index is obtained. The maximum value and peak frequency value are the core indicators of disturbance intensity, reflecting the most severe degree of disturbance. The weighted summation comprehensively considers the influence of voltage and frequency, avoiding judgment bias caused by a single indicator, and generating a more comprehensive quantitative value of disturbance severity. By determining the time period based on the comprehensive index exceeding a threshold, disturbances with insufficient intensity are further filtered out, ensuring that the finally determined time period is the truly effective disturbance that needs to be processed, avoiding the waste of computing power in subsequent frequency domain analysis on invalid data, and improving the accuracy of disturbance identification.

[0024] Furthermore, the step of extracting frequency domain features based on the second voltage signal sequence and the second frequency signal sequence to obtain anomaly analysis data specifically involves:

[0025] The second voltage signal sequence is subjected to a fast Fourier transform to obtain a frequency domain amplitude spectrum. Based on the frequency domain amplitude spectrum, the amplitude of each harmonic component is extracted, and the percentage of each harmonic content is calculated based on the amplitude of each harmonic component.

[0026] The second frequency signal sequence is subjected to a fast Fourier transform to extract frequency components, and the component with the largest amplitude among the frequency components is taken as the dominant frequency component.

[0027] Based on the percentage of each harmonic content and the pre-collected target power grid phase data, the harmonic feature vector is extracted, and the difference between the dominant frequency component and the preset rated power grid frequency is calculated to obtain the actual frequency offset value. The harmonic feature vector and the actual frequency offset value are then integrated to obtain a frequency domain feature set.

[0028] If the percentage of harmonic content in the harmonic feature vector exceeds a preset alarm threshold, or the actual frequency offset value exceeds the frequency deviation limit, it is determined to be a frequency anomaly event and an anomaly identifier is generated. The frequency domain feature set and the anomaly identifier are then integrated to obtain anomaly analysis data.

[0029] This invention extracts the percentage of harmonic content from voltage signals using Fast Fourier Transform (FFT) and the dominant frequency component from frequency signals using FFT. FFT is a core method for frequency domain analysis, converting time-domain signals into frequency-domain spectra. Extracting harmonic content allows for the determination of whether disturbances are harmonic pollution. The dominant frequency component reflects the main source of frequency shift, and the percentage of harmonic content and dominant frequency are core frequency domain features of disturbances. Different disturbance types correspond to different harmonic combinations, providing differentiated features for cloud-based classification. By constructing harmonic feature vectors and calculating actual frequency shift values, discrete harmonic parameters and frequency shifts are quantified into structured feature vectors, facilitating similarity comparison and clustering analysis in the cloud. Anomaly markers are generated through threshold settings, and the resulting anomaly analysis data is integrated to clearly identify frequency anomaly events. This allows the cloud-based classification module to directly focus on anomaly data without needing to determine data validity, improving cloud processing efficiency. Furthermore, the integrated anomaly analysis data includes frequency domain features and anomaly markers, providing a complete feature package for cloud-based classification.

[0030] Furthermore, the calculation of the similarity between the anomaly analysis data and each historical disturbance record in the pre-stored historical disturbance record set to obtain the first similarity is specifically as follows:

[0031] The similarity between the anomaly analysis data and each historical disturbance record is calculated sequentially using a dynamic time warping algorithm to obtain the first similarity score.

[0032] In this embodiment of the invention, the first similarity is calculated using the Dynamic Time Warping (DTW) algorithm. Power grid disturbance signals are time series data, which may have differences in length or rhythm. DTW is good at handling the nonlinear alignment of time series. Even if the lengths of two sequences are different, it can accurately calculate the similarity, avoid misjudgment due to low similarity caused by differences in time length, and ensure that the selected subset of historical disturbance records is highly correlated with the current disturbance.

[0033] Furthermore, the step of determining the type of disturbance event using a preset K-means clustering algorithm based on the subset of historical disturbance records and the anomaly analysis data specifically involves:

[0034] A difference operation is performed on the subset of historical disturbance records to extract the disturbance variation pattern and the periodicity feature of the disturbance;

[0035] Using a preset K-means clustering algorithm, the historical disturbance record subset and the anomaly analysis data are clustered and grouped according to the disturbance change pattern and the disturbance periodicity characteristics to obtain the grouping type. Based on the grouping type and the preset disturbance classification criteria, the disturbance event type is determined.

[0036] This invention extracts the dynamic and periodic features of disturbances through differential calculation. These features are key to distinguishing disturbance types and provide a basis for clustering differentiation. By using K-means clustering to group similar historical records and current data, it automatically identifies the natural categories in the data (without manual labeling), solving the problem of manually classifying massive amounts of historical data. Combined with preset classification standards, the clustering results are transformed into clear disturbance type identifiers, avoiding the inability to correspond to actual disturbance scenarios after clustering, and realizing a closed loop from data grouping to type implementation.

[0037] Furthermore, the step of obtaining a candidate node list for the disturbance source based on pre-stored target power grid topology data and the disturbance event type using a preset graph neural network specifically involves:

[0038] Obtain the electrical parameters of each node in the target power grid, and construct a weighted directed graph containing the electrical parameters based on the electrical parameters, pre-stored target power grid topology data, and the disturbance event type;

[0039] Starting from the edge node corresponding to the detected disturbance event, the weighted directed graph is propagated through a preset graph neural network to obtain a list of candidate nodes for the disturbance source.

[0040] This invention integrates power grid topology data, electrical parameters, and disturbance types to abstract the power grid into a weighted graph. This allows graph neural networks to trace the source based on electrical patterns, rather than just physical connections, thus improving the rationality of the tracing. By tracing back from edge monitoring points to obtain a list of candidate nodes, and leveraging the graph neural network's ability to process graph-structured data, the system propagates backward along the electrical connection path, automatically identifying nodes that may cause disturbances and generating a list of candidate nodes. This narrows the source location range from all nodes in the power grid to a few candidate nodes, avoiding blind investigation and improving the location speed.

[0041] Furthermore, the step of calculating the second similarity between each candidate node and the pre-stored historical source node, and determining the actual source node based on the second similarity, specifically involves:

[0042] The electrical parameters and load characteristics of each candidate node are extracted sequentially from the candidate node list. Based on the electrical parameters and load characteristics of each candidate node and the electrical parameters and load characteristics of the pre-stored historical source nodes, a cosine similarity is calculated to obtain a second similarity.

[0043] If the second similarity exceeds the preset matching threshold, then the corresponding candidate node is determined to be the actual source node.

[0044] This invention extracts electrical parameters and load features from candidate nodes and calculates cosine similarity. Electrical parameters and load features are essential characteristics of the source node. By comparing the feature vectors of candidate nodes with those of historical source nodes based on cosine similarity, the degree of matching between the two is quantified, avoiding subjective judgment of whether a node is the source. The actual source is determined by the similarity exceeding a threshold. The node that best matches the features of the historical source is accurately selected from the candidate list, avoiding misjudging nodes on the propagation path as source nodes and ensuring the accuracy of source location.

[0045] Furthermore, the step of generating disturbance event identification tags based on the real-time frequency value sequence specifically involves:

[0046] Based on the real-time frequency value sequence, the difference between each real-time frequency value and the preset grid rated frequency is calculated sequentially to obtain the real-time frequency offset value sequence.

[0047] Based on the real-time frequency offset value sequence, the offset change rate is calculated, and the historical offset change rate of the source node is obtained, so as to calculate the difference between the offset change rate and the historical offset rate.

[0048] If the difference is greater than a preset threshold, a strong monitoring mode is activated, and in the strong monitoring mode, the sampling frequency is adjusted to a preset multiple of the original frequency to obtain high sampling rate perturbation data.

[0049] Based on high-sampling-rate perturbation data, feature parameters such as perturbation amplitude, duration, and harmonic components are extracted to generate perturbation event identification labels.

[0050] This invention continuously monitors the frequency status of source nodes and compares real-time and historical change rates to verify whether the current disturbance is consistent with historical disturbance patterns, avoiding false source misjudgments due to monitoring errors and ensuring the reliability of source monitoring. A strong monitoring mode is activated when the difference exceeds a threshold, improving the sampling rate data. This strong monitoring mode can capture more detailed disturbance features, solving the problem of missing rapid transient processes in conventional sampling rates and providing high-resolution feature data for tags. By integrating source nodes, disturbance types, and feature parameters, feature parameters are extracted to generate tags containing disturbance information, providing structured data support for cloud database updates and early warning deployment, avoiding fragmented disturbance information.

[0051] Furthermore, the step of establishing a power grid disturbance event early warning monitoring point layout based on the disturbance event identification tag specifically involves:

[0052] Based on the disturbance event identification label, the early warning monitoring level of the source node is determined, and the early warning monitoring points of the source node are established according to the early warning monitoring level.

[0053] Based on the disturbance event identification tag and the target power grid topology data, the disturbance event propagation path is determined, and early warning monitoring points are established at each node of the disturbance event propagation path.

[0054] This invention determines the warning level based on the severity of the disturbance, and focuses on deploying monitoring equipment at high-level source nodes to ensure that key sources are controllable in real time and avoid wasting resources on low-risk nodes. By combining the disturbance propagation path and deploying monitoring equipment at nodes along the path, a full-chain monitoring network of "source + propagation path" is formed. Subsequent disturbances can be quickly detected through points along the propagation path, avoiding the need to detect disturbances only after they have spread, and improving response speed.

[0055] The above description is merely an overview of the technical solutions of the embodiments of the present invention. In order to better understand the technical means of the embodiments of the present invention and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0056] Figure 1 This is a computer system architecture diagram based on a cloud-edge computing collaborative computing model of the power Internet of Things, provided for an embodiment of the present invention. Detailed Implementation

[0057] 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.

[0058] Example 1:

[0059] like Figure 1 As shown in the figure, a computer system based on a cloud-edge computing collaborative computing model of the power Internet of Things is provided in an embodiment of the present invention. This system is suitable for monitoring power quality disturbance events. The computer system includes an edge-side time period determination module 101, an edge-side frequency domain analysis module 102, a cloud-side event classification module 103, a cloud-side source node determination module 104, and an edge-side monitoring module 105.

[0060] The edge-side time period determination module 101 is used to calculate the real-time disturbance intensity level based on the first voltage signal sequence and the first frequency signal sequence of the target power grid collected in real time, through a preset machine learning model, and determine the disturbance event time period based on the real-time disturbance intensity level when the real-time disturbance intensity level exceeds a preset intensity threshold.

[0061] In this embodiment, the edge-side time period determination module 101 calculates the real-time disturbance intensity level based on the first voltage signal sequence and the first frequency signal sequence of the target power grid acquired in real time, using a preset machine learning model. Specifically, the edge-side time period determination module 101 acquires the first voltage signal sequence and the first frequency signal sequence of several sampling points in the target power grid in real time through a preset sliding time window; based on the first voltage signal sequence, it sequentially calculates the difference between the real-time effective voltage value and the rated voltage value of each sampling point to obtain a voltage amplitude deviation sequence; based on the first frequency signal sequence, it sequentially calculates the difference between the real-time frequency value and the standard frequency value of each sampling point to obtain a frequency offset sequence; based on the voltage amplitude deviation sequence, the frequency offset sequence, a preset deviation anomaly threshold, and a preset offset anomaly threshold, it performs anomaly count of the number of abnormal sampling points to obtain anomaly statistics, and calculates the real-time disturbance intensity level based on the anomaly statistics using a preset neural network model.

[0062] In one specific embodiment, on the edge side, the microprocessor first continuously acquires the three-phase instantaneous voltage and frequency of the target power grid through a synchronous phasor measurement unit (PMU) at a fixed sampling rate of 1 kHz, forming a first voltage signal sequence and a first frequency signal sequence within a sliding time window (approximately 0.5 s) of length 512 points. Subsequently, the following three steps are performed on the 32-bit floating-point DSP core: First, the effective value of the first voltage signal sequence is calculated point by point, and the difference is calculated with the rated voltage of 220V stored in the EEPROM to obtain the voltage amplitude deviation sequence ΔV(n); Second, the standard power frequency of 50 Hz is subtracted point by point from the first frequency signal sequence to obtain the frequency offset sequence Δf(n); Finally, ΔV(n) and Δf(n) are compared with the preset deviation anomaly threshold ±5% and offset anomaly threshold ±0.2 Hz, and the number of sampling points exceeding the limit within the window is counted to obtain the anomaly statistics. The statistical results, along with the statistical characteristics (mean, variance, kurtosis) of ΔV(n) and Δf(n), are fed into a pre-quantized 3-layer fully connected neural network (6-dimensional input layer, 32-dimensional hidden layer, 1-dimensional output layer, ReLU activation, offline training with Adam optimizer, and Huber loss function) in TensorFlow Lite for inference. The network output is the real-time 0-1 normalized perturbation intensity level R. If R exceeds a preset threshold of 0.6, a power quality perturbation event is identified within the current window, triggering the subsequent perturbation event time period determination module.

[0063] In this embodiment, the edge-side time period determination module 101 determines the time period of the disturbance event based on the real-time disturbance intensity level. Specifically, the edge-side time period determination module 101 performs differential operation on the voltage amplitude deviation sequence of each sampling point to obtain the deviation change rate, and takes the sampling point where the deviation change rate exceeds a preset change rate threshold as the mutation point; based on the voltage amplitude deviation sequence corresponding to the mutation point, the disturbance start point and the disturbance end point are identified, and the maximum value of the voltage amplitude deviation sequence between the disturbance start point and the disturbance end point and the peak value of the frequency offset sequence are extracted; based on the real-time disturbance intensity level and the preset weighting coefficient, the maximum value and the peak value are weighted and summed to obtain a comprehensive disturbance index. If the comprehensive disturbance index is greater than the preset comprehensive disturbance threshold, the time period of the disturbance event is determined based on the disturbance start point and the disturbance end point.

[0064] In one specific embodiment, when the edge-side microprocessor determines that the real-time disturbance intensity level R ≥ 0.6, it immediately initiates the precise location process for the disturbance event time period: First, it performs a first-order backward difference operation on the obtained voltage amplitude deviation sequence ΔV(n) to obtain the deviation change rate sequence dΔV(n) / dt, and marks the sampling points where the absolute value of the change rate exceeds the preset change rate threshold of 3% / ms as abrupt change points; then, starting from the first abrupt change point, it traces backward along the time axis to the sampling point where the deviation first falls below the baseline noise band (±0.5%), which is recorded as the disturbance start point, and then searches backward to the sampling point where the deviation falls back into the baseline noise band and lasts for more than 50 ms, which is recorded as the disturbance end point. Within the closed interval between the disturbance time point and the disturbance end point, the maximum value V_max of ΔV(n) and the peak value f_max of the frequency offset sequence Δf(n) are extracted respectively, and the comprehensive disturbance index D is calculated according to the real-time disturbance intensity level R and the preset weight coefficients of 0.7 and k_f=0.3. If D exceeds the comprehensive disturbance threshold of 4.5, the closed interval from the disturbance time point to the disturbance end point is finally determined as the time period of this disturbance event. This time period information, together with the original sampling sequence, is packaged and sent to the cloud event classification module via the MQTT protocol. If D does not meet the threshold, it is regarded as an instantaneous disturbance, the data segment is discarded and monitoring continues.

[0065] The edge-side frequency domain analysis module 102 is used to adjust the width of the analysis window according to the time period of the disturbance event, so as to extract the second voltage signal sequence and the second frequency signal sequence through the analysis window, and to perform frequency domain feature extraction based on the second voltage signal sequence and the second frequency signal sequence to obtain abnormal analysis data.

[0066] In this embodiment, the edge-side frequency domain analysis module 102 performs frequency domain feature extraction based on the second voltage signal sequence and the second frequency signal sequence to obtain anomaly analysis data. Specifically, the edge-side frequency domain analysis module 102 performs a fast Fourier transform on the second voltage signal sequence to obtain a frequency domain amplitude spectrum, and extracts the amplitude of each harmonic component based on the frequency domain amplitude spectrum to calculate the percentage of each harmonic content based on the amplitude of each harmonic component; it also performs a fast Fourier transform on the second frequency signal sequence to extract frequency components, and takes the component with the largest amplitude among the frequency components as the dominant component. Frequency components; based on the percentage of each harmonic content and the pre-collected target power grid phase data, harmonic feature vectors are extracted, and the difference between the dominant frequency component and the preset rated power grid frequency is calculated to obtain the actual frequency offset value. The harmonic feature vectors and the actual frequency offset values ​​are then integrated to obtain a frequency domain feature set. If the percentage of harmonic content in the harmonic feature vector exceeds a preset alarm threshold, or the actual frequency offset value exceeds the frequency deviation limit, it is determined as a frequency anomaly event and an anomaly identifier is generated. The frequency domain feature set and the anomaly identifier are then integrated to obtain anomaly analysis data.

[0067] In one specific embodiment, after obtaining the disturbance event time period at the edge side, the edge computing unit immediately re-acquires the original three-phase voltage and frequency waveforms within that time period at a higher sampling rate of 12.8 kHz. After preprocessing with an 8th-order Chebyshev Type I anti-aliasing filter of 0.1-3 kHz, a second voltage signal sequence and a second frequency signal sequence of length N = 2048 are formed. The following frequency domain feature extraction steps are then performed:

[0068] First, a Blackman-Harris window is applied to the second voltage signal sequence, and a 2048-point basis-2 FFT is performed to obtain the frequency domain amplitude spectrum. The amplitude of each harmonic is accurately extracted at the 50Hz fundamental wave and its 2nd to 50th harmonics, and the percentage of each harmonic content is calculated as HR_i=H_i / ΣH_j×100%, thus constructing a 50-dimensional harmonic feature vector h_vec.

[0069] Furthermore, after applying a windowed FFT to the second frequency signal sequence, the maximum amplitude component in the 45-55 Hz range is taken as the dominant frequency component f_d, and its actual frequency offset Δf = f_d - 50 from the grid rated frequency of 50 Hz is calculated.

[0070] Furthermore, by combining the pre-acquired power grid phase data, the fundamental phase is extracted through a synchronous phase-locked loop, and the phase-amplitude coupling coefficients of the 5th, 7th, 11th, and 13th characteristic harmonics are calculated and concatenated to h_vec to form a 54-dimensional harmonic feature vector.

[0071] Furthermore, when any HR_i > alarm threshold 5% or |Δf| > frequency deviation limit 0.5 Hz, the 8-bit exception flag Flag = 0xFF is set; otherwise, Flag = 0x00.

[0072] Finally, the 54-dimensional harmonic feature vector, Δf, and anomaly identifier are integrated into a 56-byte "anomaly analysis data" structure, serialized using Protocol Buffers, and uploaded to the cloud event classification module via a 4G / 5G encrypted link.

[0073] The cloud-based event classification module 103 is used to calculate the similarity between the anomaly analysis data and each historical disturbance record in the pre-stored historical disturbance record set to obtain a first similarity, and to filter the historical disturbance record set based on the first similarity to obtain a subset of historical disturbance records, so as to determine the disturbance event type based on the subset of historical disturbance records and the anomaly analysis data using a preset K-means clustering algorithm.

[0074] In this embodiment, the cloud event classification module 103 calculates the similarity between the anomaly analysis data and each historical disturbance record in the pre-stored historical disturbance record set to obtain a first similarity. Specifically, the cloud event classification module 103 uses a dynamic time warping algorithm to sequentially calculate the similarity between the anomaly analysis data and each historical disturbance record to obtain a first similarity.

[0075] In one specific embodiment, after receiving the anomaly analysis data uploaded from the edge side, the cloud event classification module first unpacks it to obtain a 54-dimensional harmonic feature vector h_vec, frequency offset Δf, and anomaly identifier Flag; then, it performs similarity calculations on the 56-byte data against the "historical disturbance record set" deployed in the cloud time series database (InfluxDB cluster, which retains records for the most recent 24 months). The specific process is as follows: Each historical record is parsed into a 54-dimensional vector h_hist and a scalar Δf_hist, and both are standardized using z-scores to make their mean 0 and variance 1. Then, the Dynamic Time Warping (DTW) algorithm based on Sakoe-Chiba constraints (window width r=5) is used to calculate the shortest warped path distance d_h between the 54-dimensional harmonic vector of the current anomaly analysis data and the corresponding vector of each historical record, and the absolute difference d_f between Δf and Δf_hist is calculated. Finally, the first similarity S is defined as S=1 / (1+α·d_h+β·d_f), where α=0.85 and β=0.15 are weighting coefficients obtained by offline Bayesian optimization, and the value of S ranges from 0 to 1, with the closer to 1 indicating greater similarity. All historical records are sorted in descending order of S to form a candidate set for subsequent K-means clustering.

[0076] In this embodiment, the cloud-based event classification module 103 determines the type of disturbance event based on the subset of historical disturbance records and the anomaly analysis data using a preset K-means clustering algorithm. Specifically, the cloud-based event classification module 103 performs a difference operation on the subset of historical disturbance records to extract the disturbance change pattern and the disturbance periodicity feature; it then uses the preset K-means clustering algorithm to cluster the subset of historical disturbance records and the anomaly analysis data according to the disturbance change pattern and the disturbance periodicity feature, obtaining a grouping type, and determines the type of disturbance event based on the grouping type and a preset disturbance classification standard.

[0077] In one specific embodiment, after completing the first similarity calculation in the cloud, the system uses a similarity threshold of 0.65 as a boundary to extract the top 5% of high-similarity samples from the entire historical disturbance record set to form a "historical disturbance record subset" (typically 300-800 records), and combines the current anomaly analysis data with this subset to form a new clustering sample pool. Subsequently, the following K-means clustering process is executed to determine the type of disturbance event:

[0078] (1) Feature reprocessing: Perform first-order difference on the 54-dimensional harmonic feature vector of each record in the sample pool, extract the rate of change of adjacent harmonic content, and thus obtain a 54-dimensional "disturbance change law" vector; at the same time, perform 256-point FFT on the vector, take the four frequency points with the largest energy as "disturbance periodicity features", and finally send the 54+4=58-dimensional feature vector into the clusterer.

[0079] (2) Clusterer configuration: mini-batch K-means is used, batch size=100, maximum iterations 300 times; K value is determined to be 7 by offline Elbow method (corresponding to the 7 major power quality events specified in national standard GB / T 14549-2008: harmonic distortion, voltage sag, voltage swell, short interruption, flicker, oscillation transient, frequency shift).

[0080] (3) Type determination: After clustering is completed, take the centroid record of the cluster to which the current anomaly analysis data belongs, and the manually labeled event label is the type of the current disturbance event; if the current data falls at the boundary of two clusters (profile coefficient <0.25), then start the secondary SVM fine classifier (RBF kernel, γ=0.8, C=10) to make a second judgment to ensure that the classification confidence is ≥0.9.

[0081] (4) Output of results: The cloud will write the finally determined disturbance event type (such as “harmonic distortion - 5th dominant” or “voltage sag - depth 35%) along with the cluster number and centroid distance into the Kafka message queue for the downstream source node positioning module to call.

[0082] The cloud-based source node determination module 104 is used to obtain a list of candidate nodes for disturbance sources through a preset graph neural network based on pre-stored target power grid topology data and the disturbance event type, and to calculate the second similarity between each candidate node and the pre-stored historical source nodes, so as to determine the actual source node based on the second similarity; wherein, the historical source nodes and each candidate node are of the same disturbance event type.

[0083] In this embodiment, the cloud-based source node determination module 104 obtains a candidate node list of disturbance sources based on pre-stored target power grid topology data and the disturbance event type through a preset graph neural network. Specifically, the cloud-based source node determination module 104 obtains the electrical parameters of each node in the target power grid, and constructs a weighted directed graph containing the electrical parameters based on the electrical parameters, the pre-stored target power grid topology data, and the disturbance event type. Starting from the edge node corresponding to the detected disturbance event, the preset graph neural network is used to perform source propagation on the weighted directed graph to obtain a candidate node list of disturbance sources.

[0084] In one specific embodiment, after receiving a specific disturbance event type such as "harmonic distortion - 5th dominant", the cloud-based source node determination module first reads the pre-stored target power grid topology in CIM / XML format from PostgreSQL, and abstracts the entire network of 10kV and above buses, main transformers, distributed power sources, SVG, and load nodes into a weighted directed graph G=(V,E):

[0085] Node attributes: For each node in V, a 14-dimensional electrical parameter vector is formed by splicing together the fundamental voltage, current, active power, reactive power, node short-circuit capacity, 5th-13th harmonic impedance components, and load type weights (industrial 0.35, commercial 0.25, residential 0.15, etc.) uploaded in real time by the phasor measurement device.

[0086] Edge weights: 3D edge features are constructed using the branch resistance, reactance, power flow direction and event type-related harmonic attenuation coefficient α_h (e.g., 0.92 / km for the 5th harmonic) and normalized to [0,1].

[0087] Graph Neural Network: A 3-layer GraphSAGE (aggregation function mean-pooling, hidden dimension 64, ReLU, Dropout 0.2) was used. In the offline stage, it was trained with 120,000 labeled samples covering all 7 types of perturbations over 18 months. The loss function was negative log-likelihood, the optimizer was AdamW, the learning rate was 1e-3, and the early stopping patience was 15.

[0088] Online inference: Starting from the edge node that detected the disturbance, input G into the model that has been embedded in TensorFlow Serving, perform forward propagation, and the network aggregates neighborhood features layer by layer and outputs the probability p_i of each node becoming the source of the disturbance; set the probability threshold to 0.55, sort the nodes with p_i≥0.55 in descending order of probability to form a candidate node list (usually 3-7 in length), and cache the list along with the 14-dimensional features of the nodes in Redis for the next step of "cosine similarity matching".

[0089] In this embodiment, the cloud-based source node determination module 104 calculates the second similarity between each candidate node and the pre-stored historical source nodes to determine the actual source node based on the second similarity. Specifically, the cloud-based source node determination module 104 sequentially extracts the electrical parameters and load characteristics of each candidate node from the candidate node list, and performs cosine similarity calculation based on the electrical parameters and load characteristics of each candidate node and the electrical parameters and load characteristics of the pre-stored historical source nodes to obtain the second similarity. If the second similarity exceeds a preset matching threshold, the corresponding candidate node is determined to be the actual source node.

[0090] In one specific embodiment, after the candidate node list is generated in the cloud, the system immediately initiates the "historical source node matching" process to accurately locate the actual source of the disturbance. First, from the PostgreSQL historical source node database, only records that are completely consistent with the current disturbance event type (such as "5th harmonic dominance" or "voltage sag") and occurred no more than 180 days ago are retained, forming a subset H of historical source nodes. For each node C_i in the candidate node list, the following 10-dimensional feature vector is extracted in real time: fundamental voltage amplitude; 5th, 7th, 11th, and 13th harmonic voltage content; node short-circuit capacity S_k; current active load P; current reactive load Q; load power factor λ; load type code (industrial 1.0, commercial 0.7, residential 0.4); distributed power supply capacity percentage D_p; historical average harmonic impedance modulus Z_h; load change rate ΔL 1 second before and after the event. For each historical source node record in H, a 10-dimensional vector h_j is also generated using the same feature template. Subsequently, cosine similarity is used as the second similarity metric: all features are first normalized to [0,1] using min-max, and then sim(C_i,h_j)=(C_i·h_j) / (‖C_i‖‖h_j‖). The system calculates sim for C_i and all h_j in H in turn, and takes the maximum value sim_max; if sim_max exceeds the preset matching degree threshold of 0.82, C_i is immediately identified as the actual source node of this disturbance event, and its node ID, geographical coordinates, and electrical parameters are written to the Kafka message queue; if the sim_max of all candidate nodes is lower than 0.82, a manual review process is triggered, and the node with the highest sim is selected by default as the temporary source node, and the confidence level is marked as "low" for subsequent correction.

[0091] The edge-side monitoring module 105 is used to perform real-time frequency monitoring on the actual source node to obtain a real-time frequency value sequence, generate a disturbance event identification tag based on the real-time frequency value sequence, and establish a power grid disturbance event early warning monitoring point layout based on the disturbance event identification tag.

[0092] In this embodiment, the edge-side monitoring module 105 generates a disturbance event identification tag based on the real-time frequency value sequence. Specifically, the edge-side monitoring module 105 calculates the difference between each real-time frequency value and the preset rated frequency of the power grid according to the real-time frequency value sequence to obtain a real-time frequency offset value sequence; it calculates the offset change rate based on the real-time frequency offset value sequence and obtains the historical offset change rate of the source node to calculate the difference between the offset change rate and the historical offset rate; if the difference is greater than a preset threshold, a strong monitoring mode is activated, and in the strong monitoring mode, the sampling frequency is adjusted to a preset multiple of the original frequency to obtain high sampling rate disturbance data; based on the high sampling rate disturbance data, the disturbance amplitude, duration, and harmonic component feature parameters are extracted to generate a disturbance event identification tag.

[0093] In one specific embodiment, after confirming the actual source node, the edge-side monitoring module immediately performs continuous real-time sampling of the node's PMU frequency channel at a sampling rate of 6.4 kHz for 10 seconds, obtaining a real-time frequency value sequence f_raw(t) of 64,000 sampling points. The following steps are then performed to generate disturbance event identification tags:

[0094] (1) Preprocessing: DC and slow drift are removed by a 5th-order zero-phase Butterworth high-pass filter (cutoff 0.1Hz) to obtain f_hp(t).

[0095] (2) Real-time offset sequence: Calculate the difference between f_hp(t) and the grid rated frequency of 50 Hz point by point to obtain the real-time frequency offset value sequence Δf(t).

[0096] (3) Calculation of offset change rate: Perform 80-point (12.5 ms) sliding difference on Δf(t) to obtain the instantaneous offset change rate sequence r(t)=dΔf / dt.

[0097] (4) Historical offset rate comparison: Read the historical average offset change rate r_hist and its standard deviation σ of the same event type for the past 30 days from Redis; if |r(t)-r_hist|>1.5σ, then trigger "strong monitoring mode".

[0098] (5) Strong monitoring mode: Increase the sampling rate to 4 times the original frequency (25.6 kHz), continuously collect data for 2 seconds, and obtain high sampling rate disturbance data f_hi(t).

[0099] (6) Feature extraction: Add a 512-point Kaiser window with 10% overlap to f_hi(t), and extract features after FFT.

[0100] a. The peak value of the perturbation amplitude A=max(|FFT(f_hi)|) in the 0.5-45Hz passband;

[0101] b. Duration T = Δf(t) is the time between the first crossing of ±0.2Hz and the last return to ±0.05Hz;

[0102] c. Harmonic component characteristic parameters: Energy proportion of 2nd to 7th integer harmonics H2-H7.

[0103] (7) Tag encoding: The eight floating-point parameters A, T, and H2-H7 are packaged in IEEE-754 single precision, and then an 8-bit event type code (taken from the result of claim 6) and a 4-bit confidence code are concatenated to form a 40-byte "disturbance event identification tag". This tag is transmitted back to the cloud in real time via TLS encrypted UDP data packets and is simultaneously written to the local SQLite historical table for subsequent early warning monitoring point layout.

[0104] It should be noted that the present invention includes several edge-side monitoring modules, and the selection of which edge-side monitoring modules to activate is determined by the location of the actual disturbance source in the power grid.

[0105] In this embodiment, the edge-side monitoring module 105 establishes a layout of early warning monitoring points for power grid disturbance events based on the disturbance event identification tags. Specifically, the edge-side monitoring module 105 determines the early warning monitoring level of the source node based on the disturbance event identification tags, and establishes early warning monitoring points for the source node based on the early warning monitoring levels. Based on the disturbance event identification tags and the target power grid topology data, the disturbance event propagation path is determined, and early warning monitoring points are established at each node along the disturbance event propagation path.

[0106] In one specific embodiment, upon receiving a 40-byte disturbance event identification tag, the cloud-based early warning deployment engine immediately parses it into eight features, including event type code, disturbance amplitude A, and duration T. Based on a built-in "level mapping table," it calculates the early warning monitoring level L of the source node: if A ≥ 2 Hz or T ≥ 1.5s, then L = Level IV (highest); if 1 Hz ≤ A < 2 Hz and 0.5s ≤ T < 1.5s, then L = Level III, and so on. Subsequently, with the source node as the root node, the engine executes an improved Dijkstra algorithm on a pre-stored directed graph of the power grid topology. Edge weights are weighted according to branch impedance and harmonic attenuation coefficients. The engine searches for all downstream nodes whose electrical distance to the source node is within 3 km or a Class 2 bus, forming a disturbance propagation path set P. For each node in P: (1) Calculate the local monitoring level based on L and the node importance coefficient (scored from 0 to 1 by combining short-circuit capacity, load level, and number of sensitive users); (2) If the local monitoring level is greater than or equal to Level II, deploy a PMU monitoring unit with 6.4kHz continuous sampling, FFT real-time analysis, and MQTT feedback functions at the node; if the local monitoring level is greater than or equal to Level I and less than Level II, add a 1kHz sampling plug-in to the existing distribution automation terminal; if the local monitoring level is less than Level I, only retain the SCADA slow data interface.

[0107] Finally, the engine generates a "Power Grid Disturbance Event Early Warning Monitoring Point Layout File" in JSON format, which includes node ID, latitude and longitude, monitoring level, equipment model, communication method and data transmission cycle (Level IV 1s, Level III 5s, Level II 30s, Level I 300s). After being reviewed by the power grid dispatch EMS, the file is sent to each edge node for execution, completing the closed-loop layout.

[0108] This invention embodiment uses an edge-side time period determination module 101 to collect voltage and frequency signals in real time. A machine learning model is used to calculate the disturbance intensity level, first quantifying the disturbance intensity to filter out minor fluctuations and avoid processing invalid data at the edge, thus reducing computational power consumption. Leveraging the high transmission speed of edge devices, the time window for disturbance events is quickly identified, defining an effective analysis window for subsequent frequency domain analysis. This prevents the inclusion of large amounts of normal data that could distort features, serving as the time reference for all subsequent analyses. The edge-side frequency domain analysis module 102 dynamically adjusts the window width to ensure complete coverage of the disturbance waveform, solving the problems of "missed disturbance sampling" or "excessive noise sampling" in fixed-window scenarios. Frequency domain extraction converts the time-domain signal into quantifiable frequency-domain features, combining them with anomaly markers to generate "anomaly analysis data," providing a feature carrier for cloud-based classification and realizing the core value of real-time preliminary analysis at the edge. Through cloud... The edge event classification module 103 utilizes cloud computing power to process massive historical data. Through similarity filtering and K-means clustering, it addresses the problem of insufficient edge computing power hindering in-depth analysis. It also provides a basis for subsequent source location by accurately classifying disturbance types. The cloud-based source node determination module 104, combining power grid topology and disturbance type, uses graph neural networks for reverse tracing to quickly identify candidate nodes, avoiding blind investigation across the entire power grid. It then accurately selects the actual source from the candidates using a second similarity score, solving the problem of difficulty in locating disturbances due to migration, thereby improving the source location rate. The edge-side monitoring module 105 continuously monitors the located sources, generating identification tags to support the recording of full disturbance information. It also establishes early warning points covering the source and propagation path according to the tags, forming a closed-loop optimization. Subsequent similar disturbances can be quickly detected through these early warning points, improving response speed and realizing the value of one-time location and long-term early warning. Compared with existing technologies, this invention can improve the real-time performance and accuracy of power quality disturbance event monitoring through cloud-edge-edge computing power collaboration.

[0109] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0110] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A computer system based on a cloud-edge computing collaborative computing model of the power Internet of Things, characterized in that, This computer system, suitable for monitoring power quality disturbance events, includes an edge-side time period determination module, an edge-side frequency domain analysis module, a cloud-based event classification module, a cloud-based source node determination module, and an edge-side monitoring module. The edge-side time period determination module is used to calculate the real-time disturbance intensity level based on the first voltage signal sequence and the first frequency signal sequence of the target power grid collected in real time, through a preset machine learning model. When the real-time disturbance intensity level exceeds a preset intensity threshold, the disturbance event time period is determined based on the real-time disturbance intensity level. The edge-side frequency domain analysis module is used to adjust the width of the analysis window according to the time period of the disturbance event, so as to extract the second voltage signal sequence and the second frequency signal sequence through the analysis window, and to perform frequency domain feature extraction based on the second voltage signal sequence and the second frequency signal sequence to obtain anomaly analysis data. The cloud-based event classification module is used to calculate the similarity between the anomaly analysis data and each historical disturbance record in the pre-stored historical disturbance record set to obtain a first similarity. Based on the first similarity, the historical disturbance record set is filtered to obtain a subset of historical disturbance records. Then, the disturbance event type is determined based on the subset of historical disturbance records and the anomaly analysis data using a preset K-means clustering algorithm. The cloud-based source node determination module is used to obtain a list of candidate nodes for disturbance sources through a preset graph neural network based on pre-stored target power grid topology data and the disturbance event type, and to calculate the second similarity between each candidate node and the pre-stored historical source nodes, so as to determine the actual source node based on the second similarity; wherein, the historical source nodes and each candidate node are of the same disturbance event type; The edge-side monitoring module is used to perform real-time frequency monitoring on the actual source node to obtain a real-time frequency value sequence, generate disturbance event identification tags based on the real-time frequency value sequence, and establish a power grid disturbance event early warning monitoring point layout based on the disturbance event identification tags.

2. The computer system based on the cloud-edge computing power collaborative computing model of the power Internet of Things as described in claim 1, characterized in that, The step of calculating the real-time disturbance intensity level based on the first voltage signal sequence and the first frequency signal sequence of the target power grid acquired in real time, using a preset machine learning model, specifically involves: The first voltage signal sequence and the first frequency signal sequence of several sampling points in the target power grid are acquired in real time through a preset sliding time window. Based on the first voltage signal sequence, the difference between the real-time effective voltage value and the rated voltage value at each sampling point is calculated sequentially to obtain the voltage amplitude deviation sequence; Based on the first frequency signal sequence, the difference between the real-time frequency value and the standard frequency value at each sampling point is calculated sequentially to obtain the frequency offset sequence; Based on the voltage amplitude deviation sequence, frequency offset sequence, preset deviation anomaly threshold, and preset offset anomaly threshold, the number of abnormal sampling points is counted to obtain anomaly statistics. Then, based on the anomaly statistics, the real-time disturbance intensity level is calculated using a preset neural network model.

3. The computer system based on the cloud-edge computing power collaborative computing model of the power Internet of Things as described in claim 2, characterized in that, The step of determining the time period of the disturbance event based on the real-time disturbance intensity level is as follows: Perform a differential operation on the voltage amplitude deviation sequence of each sampling point to obtain the deviation change rate, and take the sampling point where the deviation change rate exceeds the preset change rate threshold as the abrupt change point; Based on the voltage amplitude deviation sequence corresponding to the mutation point, the disturbance start point and disturbance end point are identified, and the maximum value of the voltage amplitude deviation sequence between the disturbance start point and the disturbance end point and the peak value of the frequency offset sequence are extracted. Based on the real-time disturbance intensity level and the preset weighting coefficient, the maximum value and the peak value are weighted and summed to obtain a comprehensive disturbance index. If the comprehensive disturbance index is greater than the preset comprehensive disturbance threshold, the disturbance event time period is determined based on the disturbance start point and the disturbance end point.

4. The computer system based on the cloud-edge computing power collaborative computing model of the power Internet of Things as described in claim 1, characterized in that, The step of extracting frequency domain features based on the second voltage signal sequence and the second frequency signal sequence to obtain anomaly analysis data specifically involves: The second voltage signal sequence is subjected to a fast Fourier transform to obtain a frequency domain amplitude spectrum. Based on the frequency domain amplitude spectrum, the amplitude of each harmonic component is extracted, and the percentage of each harmonic content is calculated based on the amplitude of each harmonic component. The second frequency signal sequence is subjected to a fast Fourier transform to extract frequency components, and the component with the largest amplitude among the frequency components is taken as the dominant frequency component. Based on the percentage of each harmonic content and the pre-collected target power grid phase data, the harmonic feature vector is extracted, and the difference between the dominant frequency component and the preset rated power grid frequency is calculated to obtain the actual frequency offset value. The harmonic feature vector and the actual frequency offset value are then integrated to obtain a frequency domain feature set. If the percentage of harmonic content in the harmonic feature vector exceeds a preset alarm threshold, or the actual frequency offset value exceeds the frequency deviation limit, it is determined to be a frequency anomaly event and an anomaly identifier is generated. The frequency domain feature set and the anomaly identifier are then integrated to obtain anomaly analysis data.

5. The computer system based on the cloud-edge computing power collaborative computing model of the power Internet of Things as described in claim 1, characterized in that, The calculation of the similarity between the anomaly analysis data and each historical disturbance record in the pre-stored historical disturbance record set to obtain the first similarity is as follows: The similarity between the anomaly analysis data and each historical disturbance record is calculated sequentially using a dynamic time warping algorithm to obtain the first similarity score.

6. The computer system based on the cloud-edge computing power collaborative computing model of the power Internet of Things as described in claim 1, characterized in that, The step of determining the type of disturbance event using a preset K-means clustering algorithm, based on the subset of historical disturbance records and the anomaly analysis data, specifically involves: A difference operation is performed on the subset of historical disturbance records to extract the disturbance variation pattern and the periodicity feature of the disturbance; Using a preset K-means clustering algorithm, the historical disturbance record subset and the anomaly analysis data are clustered and grouped according to the disturbance change pattern and the disturbance periodicity characteristics to obtain the grouping type. Based on the grouping type and the preset disturbance classification criteria, the disturbance event type is determined.

7. The computer system based on the cloud-edge computing power collaborative computing model of the power Internet of Things as described in claim 1, characterized in that, The step of obtaining a candidate node list for the disturbance source based on pre-stored target power grid topology data and the disturbance event type using a preset graph neural network is as follows: Obtain the electrical parameters of each node in the target power grid, and construct a weighted directed graph containing the electrical parameters based on the electrical parameters, pre-stored target power grid topology data, and the disturbance event type; Starting from the edge node corresponding to the detected disturbance event, the weighted directed graph is propagated through a preset graph neural network to obtain a list of candidate nodes for the disturbance source.

8. The computer system based on the cloud-edge computing power collaborative computing model of the power Internet of Things as described in claim 1, characterized in that, The step of calculating the second similarity between each candidate node and the pre-stored historical source node, and determining the actual source node based on the second similarity, specifically involves: The electrical parameters and load characteristics of each candidate node are extracted sequentially from the candidate node list. Based on the electrical parameters and load characteristics of each candidate node and the electrical parameters and load characteristics of the pre-stored historical source nodes, a cosine similarity is calculated to obtain a second similarity. If the second similarity exceeds the preset matching threshold, then the corresponding candidate node is determined to be the actual source node.

9. A computer system based on a cloud-edge computing collaborative computing model of the power Internet of Things as described in claim 1, characterized in that, The step of generating disturbance event identification tags based on the real-time frequency value sequence specifically involves: Based on the real-time frequency value sequence, the difference between each real-time frequency value and the preset grid rated frequency is calculated sequentially to obtain the real-time frequency offset value sequence. Based on the real-time frequency offset value sequence, the offset change rate is calculated, and the historical offset change rate of the source node is obtained, so as to calculate the difference between the offset change rate and the historical offset rate. If the difference is greater than a preset threshold, a strong monitoring mode is activated, and in the strong monitoring mode, the sampling frequency is adjusted to a preset multiple of the original frequency to obtain high sampling rate perturbation data. Based on high-sampling-rate perturbation data, feature parameters such as perturbation amplitude, duration, and harmonic components are extracted to generate perturbation event identification labels.

10. The computer system based on the cloud-edge computing power collaborative computing model of the power Internet of Things as described in claim 1, characterized in that, The step of establishing a power grid disturbance event early warning monitoring point layout based on the disturbance event identification tag is specifically as follows: Based on the disturbance event identification label, the early warning monitoring level of the source node is determined, and the early warning monitoring points of the source node are established according to the early warning monitoring level. Based on the disturbance event identification tag and the target power grid topology data, the disturbance event propagation path is determined, and early warning monitoring points are established at each node of the disturbance event propagation path.

Images