Cloud-side collaborative electric energy quality monitoring method and system

Abstract

The invention provides a cloud-edge collaborative power quality monitoring method and system. The method comprises the following steps: S1, edge acquisition and preprocessing; s2, performing anomaly detection and event triggering; s3, cloud modeling and recognition: receiving abnormal event data from a plurality of edge nodes, and constructing a space-time diagram model representing the association relationship of each monitoring node in combination with the topological structure and historical data of the industrial park power distribution network; identifying the type, the occurrence position, the influence range and the severity of the power quality abnormal event, and outputting corresponding alarm information; s4, model updating and alarm interpretation: performing model adaptive updating based on data of the abnormal event, performing online training optimization on the power quality anomaly detection model, and issuing updated model parameters to each edge node to improve subsequent detection performance; and carrying out interpretation generation on the alarm information, and generating an interpretable alarm report oriented to operation and maintenance personnel. According to the invention, efficient fusion analysis and intelligent alarm of the multi-source electric energy quality data can be realized.

Description

Technical Field

[0001] This invention relates to the field of power quality monitoring and control technology, and more specifically, to a cloud-edge collaborative power quality monitoring method and system. Background Technology

[0002] Industrial park power distribution networks are widely equipped with various electrical devices and power electronic equipment, which frequently cause complex power quality problems during operation. For example, different types of loads may cause disturbances such as voltage dips, current surges, and frequency deviations, all of which adversely affect the safe and stable operation of the power grid and the quality of electricity for users. Traditional power quality monitoring methods typically rely on a single monitoring point or a limited number of data sources, making it difficult to comprehensively capture the characteristics of multi-source, heterogeneous power quality disturbances within industrial parks. A single data source cannot reflect the overall condition of the entire system, often resulting in some locally occurring power quality events going undetected in a timely manner.

[0003] Meanwhile, in terms of data processing, traditional methods have limited preprocessing capabilities for raw monitoring data, making it difficult to effectively address differences in data formats output by different acquisition devices and various measurement noise interferences. This leads to poor data consistency and false alarms and missed alarms, affecting monitoring accuracy. Furthermore, most traditional monitoring systems employ a fixed sampling rate and continuous uploading model, constantly transmitting large amounts of raw data to the central server, consuming valuable communication bandwidth and increasing the burden of data storage and computation. With the surge in the number of IoT devices in industrial parks, this "full cloud" model is unsustainable in terms of bandwidth and latency. Research shows that by filtering and analyzing data at the edge, uploading only abnormal information to the cloud can reduce cloud traffic by more than 50%. However, existing solutions often lack effective edge-cloud collaboration mechanisms and cannot flexibly adjust data upload strategies as needed, resulting in either missing important information or wasting significant bandwidth.

[0004] In terms of intelligent analysis and alarms, traditional methods are typically based on preset thresholds or simple statistical analysis, which are insufficient in identifying complex power quality disturbance patterns, resulting in low anomaly detection rates and high false alarm rates. Furthermore, most existing systems' algorithm models lack continuous learning capabilities after deployment, failing to adaptively update according to constantly changing loads and environmental conditions in industrial settings. Consequently, the sensitivity and accuracy of monitoring systems decline over time. On the other hand, many current power quality monitoring devices only provide simple alarm signals without in-depth analysis of the causes and scope of impact of anomalies, making it difficult for maintenance personnel to understand the ins and outs of events in a timely manner, thus limiting their ability to respond quickly and handle situations accurately.

[0005] In recent years, the application of cloud computing and edge computing in power systems has provided new solutions to the aforementioned problems. Cloud platforms possess powerful storage and computing capabilities, making them suitable for global modeling and complex analysis; edge computing, being close to the field, offers advantages such as low latency and high real-time performance, making it suitable for rapid local response. Effectively combining the cloud and edge can leverage their complementary strengths and play a crucial role in power quality monitoring. However, current cloud-edge collaborative monitoring solutions are still in their early stages. How to fully utilize cloud resources while ensuring real-time performance, and how to distribute computation at the edge to reduce bandwidth consumption, remain pressing technical challenges. Furthermore, the rise of distributed intelligent algorithms such as federated learning provides a new approach for collaboratively training models in power systems without centralized raw data, but it has not yet been fully applied in the field of power quality monitoring. How to utilize federated learning or other incremental learning methods to improve the model's generalization ability across different campuses and equipment scenarios while ensuring data privacy and security are also shortcomings in existing technologies.

[0006] In summary, existing power quality monitoring methods for industrial park distribution networks have significant shortcomings in areas such as multi-source data fusion, efficient anomaly detection, bandwidth optimization, adaptive model updates, and result interpretability, failing to meet the needs of refined operation and maintenance in modern smart parks. Therefore, a new technical solution is urgently needed that integrates a cloud-edge collaborative architecture and advanced intelligent algorithms to achieve end-to-end online monitoring and intelligent alarming of power quality in industrial parks, thereby improving anomaly detection rates, reducing communication overhead, and shortening response times. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings and deficiencies of the existing technology and provide a cloud-edge collaborative power quality monitoring method and system. This invention can realize efficient fusion analysis and intelligent alarm of multi-source power quality data, thereby significantly improving the accuracy, real-time performance, interpretability of results and network resource utilization of monitoring, and effectively ensuring the power supply quality and power safety of users.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solution: a cloud-edge collaborative power quality monitoring method, characterized by comprising the following steps: S1. Edge Acquisition and Preprocessing: Collect raw power quality data from multiple sources at multiple edge monitoring nodes of the industrial park power distribution network, perform real-time preprocessing on the raw data, extract abnormal features from the preprocessed multi-dimensional features, and generate preliminary abnormal indication information. S2. Anomaly Detection and Event Triggering: Based on the preliminary anomaly indication information, real-time anomaly change detection is performed at the edge monitoring node, and the monitoring data and extracted anomaly features for the relevant time period are sent to the cloud platform via the network. S3. Cloud Modeling and Recognition: The cloud platform receives abnormal event data from multiple edge nodes, combines the topology and historical data of the industrial park's power distribution network, and constructs a spatiotemporal graph model representing the correlation between each monitoring node; on the spatiotemporal graph model, it identifies the type, location, scope of impact, and severity of power quality abnormal events, and outputs corresponding alarm information. S4. Model Update and Alarm Interpretation: Based on the data of the abnormal event, the model is adaptively updated in the cloud. An incremental learning algorithm is used to train and optimize the power quality anomaly detection model online. The updated model parameters are distributed to each edge node to improve subsequent detection performance. At the same time, the alarm information is interpreted and generated, including identifying the main factors or key equipment that caused the anomaly through counterfactual reasoning analysis, and generating an interpretable alarm report for operation and maintenance personnel.

[0009] In step S1, real-time preprocessing of the raw data refers to: using the wavelet packet decomposition algorithm (WPD) to decompose the raw power quality waveform signal into multiple frequency band subspaces, filtering out noise in each frequency band and extracting transient disturbance features, thereby achieving filtering and noise reduction and signal feature extraction for signals in different frequency bands.

[0010] In step S1, extracting anomalous features from the preprocessed multidimensional features and generating preliminary anomalous indication information means: using the online robust principal component analysis model oRPCA to divide the decomposed signal features into a low-rank background part and a sparse anomalous part, so as to extract anomalous features and generate preliminary anomalous indication information.

[0011] In step S2, the Bayesian online change point detection algorithm BOCPD is used to calculate the posterior probability of abrupt changes in the power quality feature sequence, i.e., abnormal features, in real time. When the posterior probability exceeds a preset threshold, an abnormal power quality event is determined to have occurred. The edge node triggers an event reporting mechanism to send the monitoring data and extracted abnormal features for the relevant time period to the cloud platform via the network.

[0012] In step S3, the spatiotemporal graph model is based on the electrical topology of the industrial park power distribution network, with each monitoring node regarded as a graph node and the electrical connection relationship between nodes as the graph edge; the edge weight is determined by impedance distance or historical correlation. A pre-trained graph attention network-gated recurrent unit combined model (GAT+GRU) is used to jointly learn multi-source time series data on the spatiotemporal graph model to extract the spatial correlation and temporal dependence features of the distribution network monitoring data, identify the type, cause and impact range of the abnormal events, and output corresponding alarm information and analysis results.

[0013] In step S3, the alarm information includes an interpretable anomaly cause analysis; the anomaly cause analysis is as follows: the perturbation reasoning method is used on the output of the graph attention network-gated recurrent unit combined model GAT+GRU to identify the main influencing factors or key nodes causing power quality anomalies, and the anomaly cause and suggested measures are given in a human-machine readable report form.

[0014] In step S4, the online training and optimization of the power quality anomaly detection model using the incremental learning algorithm refers to using newly acquired anomaly data to fine-tune the model parameters or perform sliding window training, in order to replace the federated learning mechanism and achieve adaptive model updates.

[0015] A cloud-edge collaborative power quality monitoring system, characterized in that it includes: The edge acquisition and preprocessing module is used to acquire multi-source power quality raw data at multiple edge monitoring nodes of the industrial park power distribution network, perform real-time preprocessing on the raw data, extract abnormal features from the preprocessed multi-dimensional features, and generate preliminary abnormal indication information. The anomaly detection and event triggering module is used to perform real-time anomaly change detection on the edge monitoring node based on the preliminary anomaly indication information, and send the monitoring data and extracted anomaly features for the relevant time period to the cloud platform via the network; The cloud-based modeling and identification module receives abnormal event data from multiple edge nodes, combines the topology and historical data of the industrial park's power distribution network, and constructs a spatiotemporal graph model representing the correlation between each monitoring node. On the spatiotemporal graph model, it identifies the type, location, scope of impact, and severity of power quality anomalies and outputs corresponding alarm information. Model update and alarm interpretation are used to perform adaptive model updates. An incremental learning algorithm is used to train and optimize the power quality anomaly detection model online. The updated model parameters are then distributed to each edge node to improve subsequent detection performance. At the same time, the alarm information is interpreted and generated, including identifying the main factors or key equipment causing the anomaly through counterfactual reasoning analysis, and generating an interpretable alarm report for operation and maintenance personnel.

[0016] A cloud-edge collaborative power quality monitoring platform includes a processor and a memory for storing processor-executable programs, characterized in that: when the processor executes the program stored in the memory, it implements the aforementioned cloud-edge collaborative power quality monitoring method.

[0017] A storage medium storing a program, characterized in that: when the program is executed by a processor, it implements the above-described cloud-edge collaborative power quality monitoring method.

[0018] This invention achieves end-to-end online monitoring of power quality in industrial park distribution networks through cloud-edge collaboration, encompassing data acquisition, anomaly detection, intelligent diagnosis, and alarm interpretation. It offers several advantages: First, multi-source monitoring data undergoes fusion preprocessing and intelligent feature extraction at the edge, significantly improving the sensitivity and accuracy of anomaly detection. Second, the event-driven hierarchical reporting mechanism significantly reduces the transmission of irrelevant data, and the adaptive data scheduling strategy effectively reduces data transmission volume and cloud load by over 50%. Third, the spatiotemporal graph model in the cloud integrates information from various monitoring points, enabling more comprehensive identification and localization of complex disturbances and providing interpretable analysis of anomaly causes. Furthermore, incremental learning enables continuous self-updating of the model, allowing the system to adapt to changes in time and on-site conditions, maintaining consistently high detection performance.

[0019] In summary, this invention significantly improves the anomaly detection rate and diagnostic accuracy of power quality monitoring in industrial parks, enhances communication bandwidth utilization efficiency, shortens average alarm delay, and provides power operation and maintenance personnel with a set of intelligent, efficient, and reliable technical tools, thereby effectively ensuring the power supply quality and electricity safety of users.

[0020] Compared with the prior art, the present invention has the following advantages and beneficial effects: The cloud-edge collaborative power quality monitoring method and system of the present invention realizes efficient fusion analysis and intelligent alarm of multi-source power quality data, thereby significantly improving the accuracy, real-time performance, interpretability of results and network resource utilization of monitoring, and effectively ensuring the power supply quality and power safety of users. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall architecture of the system described in this invention, illustrating a cloud-edge collaborative power quality monitoring system, including multiple edge nodes and a cloud platform, as well as a data transmission and model distribution mechanism; Figure 2 This is a flowchart of the algorithm processing method described in this invention, showing the entire process from edge-side signal decomposition and anomaly extraction to cloud-based spatiotemporal analysis, bandwidth scheduling, and federated learning updates. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0023] Example 1 like Figure 1 and Figure 2 As shown, the cloud-edge collaborative power quality monitoring method of the present invention includes the following steps: S1. Edge Acquisition and Preprocessing: Collect raw power quality data from multiple sources at multiple edge monitoring nodes of the industrial park power distribution network, perform real-time preprocessing on the raw data, extract abnormal features from the preprocessed multidimensional features, and generate preliminary abnormal indication information.

[0024] The real-time preprocessing of the raw data refers to: using the wavelet packet decomposition algorithm (WPD) to decompose the raw power quality waveform signal into multiple frequency band subspaces, filtering out noise in each frequency band and extracting transient disturbance features, thereby achieving filtering and noise reduction and signal feature extraction for signals in different frequency bands.

[0025] Extracting anomalous features from preprocessed multidimensional features and generating preliminary anomalous indication information means using the online robust principal component analysis model oRPCA to divide the decomposed signal features into a low-rank background part and a sparse anomalous part in order to extract anomalous features and generate preliminary anomalous indication information.

[0026] Specifically, step S1 involves acquiring raw power quality data from multiple sources using edge monitoring devices deployed throughout the industrial park's power distribution network. The raw data undergoes preprocessing such as format standardization and noise suppression. Preferably, the wavelet packet decomposition algorithm (WPD) is used to decompose power signals such as voltage and current into multi-level sub-frequency bands. Different frequency bands are filtered separately to extract the characteristic energy of each band, effectively removing background noise and highlighting transient disturbance characteristics. The wavelet packet decomposition algorithm (WPD) achieves multi-scale decomposition through a recursive filter bank, which can be expressed by the following formula:

[0027]

[0028] Among them, c j,k Let represent the coefficient sequence of the k-th subband at level j, and h[m] and g[m] be the low-pass and high-pass filter coefficients of wavelet packet decomposition, respectively. By simultaneously performing the above filtering and decimation on all subband signals at each level, finer-grained subband components for the next level can be obtained. After WPD multi-scale decomposition, the noise and irrelevant components in the original waveform are filtered out step by step, and the power quality disturbance characteristics of each frequency band are highlighted.

[0029] Subsequently, the online robust principal component analysis (oRPCA) model is used to perform online low-rank decomposition on the signal features after WPD decomposition. This separates the low-rank background component (reflecting stable components under normal operating conditions) from the sparse anomaly component (reflecting transient anomalies), enabling preliminary extraction of abnormal disturbances and generating power quality anomaly features. For example, under normal conditions, the voltage waveform mainly consists of the power frequency fundamental component (which can be considered as the low-rank background component); when anomalies such as voltage drops occur, short-term abrupt changes appear in the waveform (reflected as the sparse anomaly component). The online robust principal component analysis (oRPCA) model can quickly detect these sparse anomalies and dynamically update the background model, thereby obtaining the anomaly indication value sequence at the current moment. This invention can decompose the online robust principal component analysis (oRPCA) model into solving for the low-rank matrix L and the sparse matrix S of the data matrix X:

[0030] in Represents a low-rank matrix L nuclear norm number, Representing a sparse matrix S The element norm is given by λ, which is a parameter that balances the low-rank and sparse terms. By optimizing the objective function online, the online robust principal component analysis model oRPCA can progressively update the low-rank matrix. L (Reflecting the normal state) and sparse matrix S (Reflects abnormal disturbances) and enables real-time extraction of abnormal parts in the data stream.

[0031] S2. Anomaly Detection and Event Triggering: Based on the preliminary anomaly indication information, real-time anomaly change detection is performed at the edge monitoring node, and the monitoring data and extracted anomaly features for the relevant time period are sent to the cloud platform via the network.

[0032] Step S2 involves using the Bayesian online change point detection algorithm (BOCPD) to calculate the posterior probability of abrupt changes in the power quality feature sequence, i.e., abnormal features, in real time. When the posterior probability exceeds a preset threshold, an abnormal power quality event is determined to have occurred. The edge node triggers an event reporting mechanism to send the monitoring data and extracted abnormal features for the relevant time period to the cloud platform via the network.

[0033] Specifically, step S2 involves real-time change detection at the edge based on the aforementioned anomaly indicator value sequence to determine when the anomalous disturbance occurred and whether it reached a level requiring reporting. Preferably, the edge node runs the Bayesian Online Change Detection Algorithm (BOCPD) to evaluate the statistical characteristics of the anomaly indicator values ​​online. The BOCPD algorithm calculates the posterior probability of a change at the current time when new data arrives. (where r) tr represents the runtime ending at time t. t =0 indicates that a turning point has occurred at this point. This posterior probability can be calculated recursively, for example:

[0034] Where H is the prior hazard function (probability of the change point). When A sharp increase compared to the previous moment indicates a significant anomaly has just occurred, at which point the edge node immediately detects a power quality anomaly. Simultaneously, to accommodate disturbances of varying intensities, the edge node can set an anomaly intensity threshold; when the anomaly indication value exceeds this threshold, an event report is also triggered. Through this real-time detection mechanism, edge nodes can typically detect anomalies and respond within tens of milliseconds, significantly reducing response time compared to traditional methods relying on centralized analysis by a central server. Once the edge node determines a power quality anomaly has occurred, it immediately reports the relevant information to the cloud platform via the network. The reported information includes the time of the event, its duration, key characteristic indicators (such as the percentage voltage drop), and locally extracted anomaly feature data. Compared to the complete raw waveform, this information is small in volume and consumes almost no bandwidth, but it provides ample clues for further in-depth analysis in the cloud.

[0035] S3. Cloud Modeling and Recognition: The cloud platform receives abnormal event data from multiple edge nodes, combines the topology and historical data of the industrial park's power distribution network, and constructs a spatiotemporal graph model representing the correlation between each monitoring node; on the spatiotemporal graph model, it identifies the type, location, scope of impact, and severity of power quality abnormal events, and outputs corresponding alarm information.

[0036] The spatiotemporal graph model is based on the electrical topology of the industrial park power distribution network. Each monitoring node is regarded as a graph node, and the electrical connection relationship between nodes is regarded as the graph edge. The edge weight is determined by impedance distance or historical correlation. The pre-trained graph attention network-gated cyclic unit combination model GAT+GRU is used to jointly learn the multi-source time series data on the spatiotemporal graph model to extract the spatial correlation and time dependence features of the power distribution network monitoring data, identify the type, cause and impact range of the abnormal events, and output the corresponding alarm information and analysis results.

[0037] The above alarm information includes an interpretable analysis of the cause of the anomaly; the analysis of the cause of the anomaly is as follows: the perturbation reasoning method is used on the output of the graph attention network-gated recurrent unit combined model GAT+GRU to identify the main influencing factors or key nodes that cause power quality anomalies, and the cause of the anomaly and suggested measures are given in a human-machine readable report.

[0038] Specifically, step S3 involves the cloud platform, after receiving reports of abnormal events from multiple edge nodes, first combining the network topology and historical operational data of the park's power distribution network, starting from the monitoring node involved in the event, extracting a time window of length T for the node and its neighboring nodes before and after the event, forming a spatiotemporal data block. (Where N is the number of nodes and F is the feature dimension of each node). For example, setting the electrical connection or geographical proximity relationship of the park to construct a graph G=(V, E), the weighted adjacency (or edge weight) matrix A=[w] is obtained according to impedance distance / historical correlation. ij Based on this, the pre-trained graph attention network-gated recurrent unit combined model GAT+GRU is invoked to perform spatiotemporal joint encoding and discrimination of X. The specific data transfer is as follows: (1) Time-series spatial attention encoding (GAT) For each moment within the time window Node feature matrix Apply graph attention independently to obtain the spatial encoding at that moment. Let the input of the i-th node at time t be... Through linear mapping:

[0039] in, The node features are those obtained after linear mapping; Let F be the input feature vector of node i at time t; Given a learnable linear transformation matrix, the F-dimensional features are mapped to... Dimension. After obtaining the linear mapping, for the neighbors of i. Calculate the unnormalized attention score:

[0040] in, Scoring (scalar) for unnormalized attention. Let be a linear rectified function with leakage, and 'a' be the attention parameter vector. For the inner product operation of vector a ( (indicates transpose) This represents concatenating the mapping features of the two nodes along the feature dimension. After scoring, attention weights are obtained by softmax normalization using the adjacency mask.

[0041] in, These are the normalized attention weights; This is the natural exponential function, used to convert scores into positive weights; The denominator represents the normalized result after scoring and exponentiation of all neighbors of node i. Let be the set of neighbors of node i. Finally, the neighborhood information is weighted and aggregated to obtain the spatial code at time t:

[0042] in, The output of GAT represents the spatial aggregation representation of node i at time t, which serves as the input for subsequent time modeling. It is an element-wise nonlinear activation function; For all neighboring nodes of node i at time t... Features By attention weight Perform a weighted summation to obtain the comprehensive information about the neighbors of node i; For the real number field A 3D vector space. Stacking all nodes... .

[0043] (2) Node-by-node temporal recursive coding (GRU) For each node i, its spatial encoding sequence over the entire time window is determined. , as the input sequence of the GRU (i.e., let The hidden state is updated recursively over time to extract time-dependent features:

[0044] Where z t To update the gate vector, r t To reset the gate vector, h is the candidate hidden state vector. t Let be the hidden state vector at the current time step. The sigmoid function represents the sigmoid function, which maps real numbers to [0,1]. t Let W be the input feature vector at the current time. Z U Z Generate z respectively t The weight matrix, h t-1 Let W be the hidden state vector from the previous time step. r U r Don't generate r t The weight matrix, W h U h Generate respectively The weight matrix, Represents the hyperbolic tangent function, which maps real numbers to [-1, 1]. This is an element-wise multiplication operation. The final value is taken at the last time step. (Or, using time pooling) as the temporal feature representation of node i. After aggregating all nodes, we obtain... .

[0045] (3) Judgment and output exist The task header above completes event recognition and localization: First, it performs global averaging / attention reading and outputs the event type and confidence score after fully connected processing and softmax; second, it performs processing on each node... The probability of being affected is obtained through a sigmoid function and used for source node / affect range estimation. The model also retains the attention weights of GAT. Using GRU time contribution as an interpretable criterion, for example, the high attention synchronization change between node 1 and node 2 near t0 represents a common source disturbance. After the above process, the cloud model identifies the type, possible source, and affected area of ​​the current anomaly, and distributes policy / parameter updates to the edge as needed to complete the loop. The cloud analysis can determine: "This event is a localized voltage dip, caused by the start-up of a large motor in the park, leading to a drop in bus voltage. Its characteristic is a 15% instantaneous voltage drop within 0.2 seconds, followed by a return to normal within approximately 2 seconds." The model also provides a list of the main affected nodes and the possible locations of disturbance sources. If the model has multiple possible outcomes, it can also provide the confidence levels of each for maintenance personnel's reference.

[0046] After anomaly identification is completed, the cloud-based interpretable output module generates alarm information for maintenance personnel based on the identification results and pushes it to the monitoring center. The alarm information includes: "Alarm: Voltage sag event. Occurrence time: 10:30:05, Duration: 0.3s. Affected areas: Distribution room A and B busbars. Possible cause: Impulsive load startup in workshop 1 (based on disturbance reasoning analysis). Recommended measures: Check the start-up and shutdown strategies of high-power equipment in workshop 1, and activate reactive power compensation devices as appropriate." This type of information transforms the originally obscure power monitoring data into clear text descriptions, facilitating maintenance personnel to quickly locate problems and take corrective action.

[0047] The aforementioned explainable mechanism, based on the counterfactual reasoning algorithm of this invention, identifies the key factors most influential on anomaly detection by applying small perturbations to the model's input features and observing changes in the model's output. Essentially, this mechanism solves for the minimum input change required to transform an anomaly detection into a normal one, thereby determining the root cause of the anomaly. It can be formalized as the following optimization problem, with the minimum perturbation being:

[0048] This allows F(x+δ) to be classified as a normal operating condition. Here, x is the feature vector during the abnormal condition, and δ is the artificially constructed input perturbation. Indicates the "magnitude" of the disturbance. The system outputs the operating condition judgment result as a discriminant function or classification model. Once the minimum disturbance is found, the features corresponding to its non-zero components are the key factors that need to be changed to eliminate the anomaly. Based on this, the system identifies the dominant factors causing the anomaly and provides explanations.

[0049] S4. Model Update and Alarm Interpretation: Based on the data of the abnormal event, the model is adaptively updated in the cloud. An incremental learning algorithm is used to train and optimize the power quality anomaly detection model online. The updated model parameters are distributed to each edge node to improve subsequent detection performance. At the same time, the alarm information is interpreted and generated, including identifying the main factors or key equipment that caused the anomaly through counterfactual reasoning analysis, and generating an interpretable alarm report for operation and maintenance personnel.

[0050] The above-mentioned use of incremental learning algorithm to train and optimize the power quality anomaly detection model online refers to using newly acquired anomaly data to fine-tune the model parameters or perform sliding window training, in order to replace the federated learning mechanism to achieve adaptive model updates.

[0051] Specifically, step S4 involves the system entering an adaptive optimization phase after the event identification and alarm generation are completed in the cloud. On one hand, the cloud can temporarily adjust the data communication strategy between the edge and the cloud based on the current impact range and severity of the event to optimize the overall utilization of monitoring resources. For example, after confirming a regional anomaly, the cloud may instruct edge nodes in the relevant area to increase the sampling rate or upload frequency of key power parameters in the short term to closely track the event's development; for other edge nodes in unrelated areas, their data upload frequency can be reduced or the upload of secondary indicators can be suspended, thus prioritizing limited communication resources for key areas. Through this mechanism, the system can dynamically optimize data scheduling during complex events, ensuring timely upload of key data while avoiding unnecessary bandwidth consumption and guaranteeing the stable operation of the entire monitoring network.

[0052] On the other hand, the cloud triggers an adaptive update process for the model, using new data from this event to further improve the performance of the anomaly detection model. The update process is as follows: The cloud incorporates data fragments related to this anomaly (including data during the anomaly and corresponding normal state data) into the model training set, and uses an incremental learning algorithm to train and optimize the global anomaly detection model online. For example, the cloud can input newly collected anomaly samples and corresponding normal samples into a predetermined deep learning model for several more rounds of training, minimizing the loss function to update the model parameters. Through continuous online fine-tuning, the model can strengthen its ability to identify the anomaly pattern. After training, the cloud distributes the updated model parameters to each edge node, replacing the old model, so that the anomaly detection model of the entire network is improved synchronously. It is worth mentioning that model updates do not need to be performed after every event; they can be triggered periodically as needed or after major anomaly events. Through long-term iterative training, the model will continuously accumulate experience and become more robust. For example, it can gradually learn to distinguish the feature patterns of different types of voltage anomalies (sags, short interruptions, etc.), improving detection accuracy. The entire incremental update process makes full use of the new information collected at the edge, and at the same time, since it mainly transmits model parameters rather than raw data, it takes into account the privacy and security of the data of each node to a certain extent.

[0053] The cloud-edge collaborative power quality monitoring system of the present invention includes: The edge acquisition and preprocessing module is used to acquire multi-source power quality raw data at multiple edge monitoring nodes of the industrial park power distribution network, perform real-time preprocessing on the raw data, extract abnormal features from the preprocessed multi-dimensional features, and generate preliminary abnormal indication information. The anomaly detection and event triggering module is used to perform real-time anomaly change detection on the edge monitoring node based on the preliminary anomaly indication information, and send the monitoring data and extracted anomaly features for the relevant time period to the cloud platform via the network; The cloud-based modeling and identification module receives abnormal event data from multiple edge nodes, combines the topology and historical data of the industrial park's power distribution network, and constructs a spatiotemporal graph model representing the correlation between each monitoring node. On the spatiotemporal graph model, it identifies the type, location, scope of impact, and severity of power quality anomalies and outputs corresponding alarm information. Model update and alarm interpretation are used to perform adaptive model updates. An incremental learning algorithm is used to train and optimize the power quality anomaly detection model online. The updated model parameters are then distributed to each edge node to improve subsequent detection performance. At the same time, the alarm information is interpreted and generated, including identifying the main factors or key equipment causing the anomaly through counterfactual reasoning analysis, and generating an interpretable alarm report for operation and maintenance personnel.

[0054] Example 2 This embodiment provides a cloud-edge collaborative power quality monitoring platform, which includes a processor and a memory for storing processor-executable programs. When the processor executes the program stored in the memory, it implements the cloud-edge collaborative power quality monitoring method of Embodiment 1.

[0055] Example 3 This embodiment provides a storage medium, which is a computer-readable storage medium that stores a computer program. When the program is executed by a processor, the processor executes the computer program stored in the memory to implement the cloud-edge collaborative power quality monitoring method of the above embodiment one.

[0056] The storage medium described in this embodiment can be a disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), USB flash drive, portable hard drive, etc.

[0057] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A cloud-edge collaborative power quality monitoring method, characterized in that: Includes the following steps: S1. Edge Acquisition and Preprocessing: Collect raw power quality data from multiple sources at multiple edge monitoring nodes of the industrial park power distribution network, perform real-time preprocessing on the raw data, extract abnormal features from the preprocessed multi-dimensional features, and generate preliminary abnormal indication information. S2. Anomaly Detection and Event Triggering: Based on the preliminary anomaly indication information, real-time anomaly change detection is performed at the edge monitoring node, and the monitoring data and extracted anomaly features for the relevant time period are sent to the cloud platform via the network. S3. Cloud-based modeling and identification: The cloud platform receives abnormal event data from multiple edge nodes, and combines the topology and historical data of the industrial park's power distribution network to construct a spatiotemporal graph model representing the correlation between each monitoring node. The spatiotemporal graph model identifies the type, location, scope of impact, and severity of power quality anomalies, and outputs corresponding alarm information. S4. Model Update and Alarm Interpretation: Based on the data of the abnormal event, the model is adaptively updated in the cloud. An incremental learning algorithm is used to train and optimize the power quality anomaly detection model online. The updated model parameters are distributed to each edge node to improve subsequent detection performance. At the same time, the alarm information is interpreted and generated, including identifying the main factors or key equipment that caused the anomaly through counterfactual reasoning analysis, and generating an interpretable alarm report for operation and maintenance personnel.

2. The cloud-edge collaborative power quality monitoring method according to claim 1, characterized in that: In step S1, real-time preprocessing of the raw data refers to: using the wavelet packet decomposition algorithm (WPD) to decompose the raw power quality waveform signal into multiple frequency band subspaces, filtering out noise in each frequency band and extracting transient disturbance features, thereby achieving filtering and noise reduction and signal feature extraction for signals in different frequency bands.

3. The cloud-edge collaborative power quality monitoring method according to claim 2, characterized in that: In step S1, extracting anomalous features from the preprocessed multidimensional features and generating preliminary anomalous indication information means: using the online robust principal component analysis model oRPCA to divide the decomposed signal features into a low-rank background part and a sparse anomalous part, so as to extract anomalous features and generate preliminary anomalous indication information.

4. The cloud-edge collaborative power quality monitoring method according to claim 1, characterized in that: In step S2, the Bayesian online change point detection algorithm BOCPD is used to calculate the posterior probability of abrupt changes in the power quality feature sequence, i.e., abnormal features, in real time. When the posterior probability exceeds a preset threshold, an abnormal power quality event is determined to have occurred. The edge node triggers an event reporting mechanism to send the monitoring data and extracted abnormal features for the relevant time period to the cloud platform via the network.

5. The cloud-edge collaborative power quality monitoring method according to claim 1, characterized in that: In step S3, the spatiotemporal graph model is based on the electrical topology of the industrial park power distribution network, with each monitoring node regarded as a graph node and the electrical connection relationship between nodes as the graph edge; the edge weight is determined by impedance distance or historical correlation. A pre-trained graph attention network-gated recurrent unit combined model (GAT+GRU) is used to jointly learn multi-source time series data on the spatiotemporal graph model to extract the spatial correlation and temporal dependence features of the distribution network monitoring data, identify the type, cause and impact range of the abnormal events, and output corresponding alarm information and analysis results.

6. The cloud-edge collaborative power quality monitoring method according to claim 5, characterized in that: In step S3, the alarm information includes an interpretable anomaly cause analysis; the anomaly cause analysis is as follows: the perturbation reasoning method is used on the output of the graph attention network-gated recurrent unit combined model GAT+GRU to identify the main influencing factors or key nodes causing power quality anomalies, and the anomaly cause and suggested measures are given in a human-machine readable report form.

7. The cloud-edge collaborative power quality monitoring method according to claim 1, characterized in that: In step S4, the online training and optimization of the power quality anomaly detection model using the incremental learning algorithm refers to using newly acquired anomaly data to fine-tune the model parameters or perform sliding window training, in order to replace the federated learning mechanism and achieve adaptive model updates.

8. A cloud-edge collaborative power quality monitoring system, characterized in that: include: The edge acquisition and preprocessing module is used to acquire multi-source power quality raw data at multiple edge monitoring nodes of the industrial park power distribution network and perform real-time preprocessing on the raw data. Extract abnormal features from the preprocessed multidimensional features and generate preliminary abnormality indication information; The anomaly detection and event triggering module is used to perform real-time anomaly change detection on the edge monitoring node based on the preliminary anomaly indication information, and send the monitoring data and extracted anomaly features for the relevant time period to the cloud platform via the network; The cloud-based modeling and identification module is used to receive abnormal event data from multiple edge nodes, and combine it with the topology and historical data of the industrial park's power distribution network to construct a spatiotemporal graph model representing the correlation between each monitoring node. The spatiotemporal graph model identifies the type, location, scope of impact, and severity of power quality anomalies, and outputs corresponding alarm information. Model update and alarm interpretation are used to perform adaptive model updates. An incremental learning algorithm is used to train and optimize the power quality anomaly detection model online. The updated model parameters are then distributed to each edge node to improve subsequent detection performance. At the same time, the alarm information is interpreted and generated, including identifying the main factors or key equipment causing the anomaly through counterfactual reasoning analysis, and generating an interpretable alarm report for operation and maintenance personnel.

9. A cloud-edge collaborative power quality monitoring platform, comprising a processor and a memory for storing processor-executable programs, characterized in that: When the processor executes the program stored in the memory, it implements the cloud-edge collaborative power quality monitoring method according to any one of claims 1-7.

10. A storage medium storing a program, characterized in that: When the program is executed by the processor, it implements the cloud-edge collaborative power quality monitoring method according to any one of claims 1-7.

Images