Active power distribution network abnormity sensing system based on edge calculation

Abstract

The invention discloses an active power distribution network anomaly sensing system based on edge calculation, and relates to the technical field of active power distribution networks, and the system comprises a hyperplane boundary module which obtains the prior knowledge of the triggering performance of each known fault event, constructs a fault event development performance causal path, trains a fault event stage development recognition model, and generates a hyperplane boundary; the fault event triggering risk probability module is used for acquiring real-time operation data, performing fitting division on the real-time operation data and a hyperplane boundary of the real-time operation data, determining that the real-time operation data points to a fault event development stage, constructing a real-time fault event triggering risk assessment model and generating a real-time operation data points to a fault event triggering risk probability; and the optimal inhibition action module is used for constructing a deep reinforcement learning controller, establishing a fault event triggering risk probability inhibition action intervention function and generating an optimal inhibition action based on each known fault event inhibition action database. According to the invention, the active defense capability of the active power distribution network to potential faults is improved.

Description

Technical Field

[0001] This invention relates to the field of active power distribution network technology, specifically to an active power distribution network anomaly sensing system based on edge computing. Background Technology

[0002] In existing technologies, anomaly detection and fault handling in active distribution networks mainly rely on centralized monitoring systems. These systems suffer from delays in data processing and decision response, making it difficult to meet the stringent real-time requirements of distributed generation resources with high penetration rates. Traditional methods often rely on static thresholds or simple rules for alarms, lacking the ability to dynamically model complex fault evolution processes and accurately identify stages, resulting in delayed fault warnings and high false alarm rates. Existing suppression strategies are usually preset, fixed action sequences that cannot be adaptively optimized and adjusted according to real-time operating status and dynamic risk assessment, often leading to overly aggressive or insufficient control actions. It is difficult to achieve the optimal balance between ensuring system safety and minimizing operational losses. Existing systems lack flexibility and self-learning capabilities when dealing with new or unknown fault modes. Summary of the Invention

[0003] To address the aforementioned technical issues, an active power distribution network anomaly sensing system based on edge computing is provided. This technical solution solves the problem that existing systems lack flexibility and self-learning capabilities when dealing with new or unknown fault modes.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: An active distribution network anomaly detection system based on edge computing includes: Hyperplane boundary module, fault event triggering risk probability module, optimal suppression action module; The fault event triggering risk probability module is electrically connected to the hyperplane boundary module, and the optimal suppression action module is electrically connected to the fault event triggering risk probability module. The hyperplane boundary module acquires prior knowledge of the known triggering behavior of each fault event of several distributed generation resource equipment types in the active distribution network, constructs the causal path of the development behavior of each known fault event of the generation resource equipment type, trains the stage development recognition model of each fault event, and generates the hyperplane boundary of the development stage of each known fault event of the generation resource equipment type. The fault event triggering risk probability module acquires real-time operating data of several distributed generation resource equipment types in the active distribution network, fits and divides the hyperplane boundary of the known fault event development stage of each generation resource equipment type, determines the real-time operating data of the generation resource equipment type pointing to the known fault event development stage, constructs a real-time fault event triggering risk assessment model, and generates the real-time operating data of several distributed generation resource equipment types in the active distribution network pointing to the known fault event triggering risk probability. The optimal suppression action module, based on the known suppression action database for each fault event of the power generation resource equipment type and the real-time operation data pointing to the known trigger risk probability of each fault event, constructs a deep reinforcement learning controller, establishes an intervention function for the suppression action of the fault event trigger risk probability, and generates the optimal suppression action pointing to the known trigger risk probability of each fault event based on the real-time operation data.

[0005] Preferably, the hyperplane boundary module specifically includes: The original time-series dataset establishment unit is based on the historical database of the active distribution network. It obtains historical operation data of the active distribution network according to several distributed generation resource equipment types, including historical SCADA data, fault recording data, equipment logs and operation and maintenance reports. It aligns timestamps and performs data preprocessing to establish the original time-series dataset of the historical operation of the active distribution network. The types of distributed power generation equipment include: photovoltaic inverters, wind power converters, and energy storage PCS. Historical SCADA data includes: time-series data of voltage, current, temperature and power in the active distribution network; The fault event set establishment unit extracts the time series of historical fault cases of the active distribution network based on the original time series dataset of the historical operation of the active distribution network. Combining expert experience manuals and operation and maintenance records, it automatically parses the text using named entity recognition technology to generate standardized fault descriptions of historical fault cases of the active distribution network and constructs a historical fault event set of the active distribution network. The causal path generation unit, for each fault event in the historical fault event set of the active distribution network, based on the physical models of several distributed generation resource equipment types and historical fault cases of the active distribution network combined with the expert group, takes each known non-occurring fault event of the generation resource equipment type as the top event of the fault tree, analyzes the known fault factors of the generation resource equipment type that affect the top event of the fault tree and calculates the probability of occurrence of the factors, establishes the fault tree of the fault event of the generation resource equipment type, and obtains the causal path of the development of each known fault event of the generation resource equipment type.

[0006] Preferably, the hyperplane boundary module also includes: The stage label unit is designed with a fixed sliding time window. Anchor point mapping is used to define the causal path of each known fault event of power generation resource equipment type in the original time series dataset of the active distribution network in the historical operation within the time window and its development performance. The data segments corresponding to each development stage are located for each fault event. Through threshold rules, the data segments are labeled with the corresponding stage labels, including normal stage, budding stage, development stage and adjacent stage. A dataset of known fault events of power generation resource equipment type based on stage labels is established. To further explain, the normal stage includes: no abnormality of any fault factor is detected; the nascent stage includes: more than one fault factor shows a slight abnormality for the first time; the development stage includes: the number of fault factors increases and the degree of abnormality intensifies, and the system state continues to deteriorate; the near stage includes: the abnormal indicators of fault factors reach the fault threshold.

[0007] Preferably, the hyperplane boundary module also includes: The time-domain feature extraction unit calculates the mean, standard deviation, skewness, kurtosis, root mean square, waveform factor, and peak factor for each data segment in the known fault event dataset based on stage labels of power generation resource equipment types within the time window, obtains the statistical features of each data segment, and extracts its time-domain features. The frequency domain feature extraction unit uses Fast Fourier Transform to convert the time domain signal of each data segment in the known fault event dataset of power generation resource equipment type based on stage label within the time window into a frequency domain signal, and extracts the frequency band energy, dominant frequency component and harmonic distortion rate of each data segment to obtain its frequency domain features. The comprehensive feature vector establishment unit calculates the average slope of the known fault event datasets of power generation resource equipment types based on stage labels within the time window, captures their trend features, combines the time domain features and frequency domain features of each data segment, performs normalization processing, and concatenates them to obtain the comprehensive feature vector of each data segment in the known fault event datasets of power generation resource equipment types based on stage labels.

[0008] Preferably, the dimensionality reduction unit uses principal component analysis algorithm to perform data-based processing on the comprehensive feature vector of each data segment in the known fault event dataset of power generation resource equipment type based on stage label, calculates the feature covariance matrix of each data segment, performs eigenvalue decomposition to obtain the feature value and feature vector of each data segment, and selects the top k feature values ​​as principal components to obtain the dimensionality-reduced comprehensive feature vector of each data segment in the known fault event dataset of power generation resource equipment type based on stage label; The development identification model training unit uses the SVM support vector machine algorithm. It takes the comprehensive feature vector of each data segment in the known fault event dataset of power generation resource equipment type based on stage label after dimensionality reduction and the corresponding stage label as input, and introduces the radial basis function as the kernel function to train the stage development identification model of each fault event and generate the hyperplane boundary of the development stage of each known fault event of power generation resource equipment type.

[0009] Preferably, the fault event triggering risk probability module specifically includes: The real-time operation data comprehensive feature vector extraction unit, based on fieldbus, collects real-time operation data of several distributed generation resource equipment types in the active distribution network, performs data preprocessing, and combines the comprehensive feature vector acquisition technology of each data segment in the known fault event dataset of each generation resource equipment type based on stage label to extract the comprehensive feature vector of the real-time operation data of several distributed generation resource equipment types in the active distribution network within the time window.

[0010] Preferably, the fault event triggering risk probability module also includes: The fault development stage segmentation unit takes the comprehensive feature vector of real-time operation data of several distributed generation resource equipment types in the active distribution network within the time window as input, associates and pairs it with the stage development identification model of each fault event, calculates the functional interval between the comprehensive feature vector of the real-time operation data and the known hyperplane boundary of each fault event development stage of the generation resource equipment type, and determines the fault development stage to which the comprehensive feature vector of the real-time operation data belongs based on the positive or negative sign of the functional interval. The highest confidence level output by each fault event stage development identification model is selected to determine the real-time operation data of the generation resource equipment type pointing to the known development stages of each fault event.

[0011] Preferably, the fault event triggering risk probability module also includes: The fault event triggering risk probability determination unit, based on the real-time operation data of the determined power generation resource equipment type pointing to the known development stages of each fault event, calculates the geometric distance from the real-time operation data of the power generation resource equipment type to the hyperplane boundary of the corresponding fault stage, performs mapping and normalization processing, calculates the fault event triggering risk probability value of the real-time operation data of the power generation resource equipment type, constructs a real-time fault event triggering risk assessment model, and generates the known fault event triggering risk probability of several distributed power generation resource equipment types in the active distribution network pointing to the real-time operation data of each fault event.

[0012] Preferably, the optimal suppression action module specifically includes: The suppression action database establishment unit integrates the specifications of power generation resource equipment manufacturers, power grid dispatching specifications and expert experience, defines the suppression actions that can be executed at each stage of the development of each known fault type, sets the expected multi-level risk control objectives triggered by each known fault event, and establishes a database of suppression actions for each known fault event of power generation resource equipment type. The executable suppression actions include: alarm, power reduction operation, control mode adjustment, switching to backup equipment, and orderly shutdown; the multi-level risk control targets include: risk probability <0.3 in the nascent stage and risk probability <0.6 in the development stage.

[0013] Preferably, the optimal suppression action module also includes: The deep reinforcement learning controller construction unit uses the comprehensive feature vector of real-time operation data of several distributed generation resource equipment types in the active distribution network within a time window, the development stage of each known fault event pointed to by the real-time operation data of each generation resource equipment type, and the trigger risk probability of each known fault event pointed to by the real-time operation data as the state space. The effective action subset that conforms to the current fault type in the known fault event suppression action database of each generation resource equipment type is used as the action space. The reduction amount of each fault event trigger risk probability, the generation penalty, and the action frequency penalty are used as the composite reward function. The local system of the active distribution network under edge node control is used as the environment. The deep reinforcement learning controller is constructed to calculate the state of the local system of the active distribution network and the fault event trigger risk probability at the next time step, establish the fault event trigger risk probability suppression action intervention function, and generate the optimal suppression action pointed to by the real-time operation data of each known fault event trigger risk probability.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes an active distribution network anomaly detection scheme based on edge computing. This scheme achieves intelligent processing of active distribution network faults from early detection and dynamic risk assessment to active suppression by constructing a collaborative working system of a hyperplane boundary module, a fault event triggering risk probability module, and an optimal suppression action module. The system can accurately identify the fault development stage of complex distributed generation resource equipment, quantitatively assess the fault triggering risk probability in real time, and adaptively generate the optimal suppression action that balances safety and economy based on deep reinforcement learning, thereby improving the active defense capability, operational reliability, and risk control refinement level of the active distribution network against potential faults. Attached Figure Description

[0015] Figure 1 This is a framework diagram of an active power distribution network anomaly sensing system based on edge computing. Detailed Implementation

[0016] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0017] Reference Figure 1 As shown, the active distribution network anomaly sensing system based on edge computing includes: Hyperplane boundary module, fault event triggering risk probability module, optimal suppression action module; The fault event triggering risk probability module is electrically connected to the hyperplane boundary module, and the optimal suppression action module is electrically connected to the fault event triggering risk probability module. The hyperplane boundary module acquires prior knowledge of the known triggering behavior of each fault event of several distributed generation resource equipment types in the active distribution network, constructs the causal path of the development behavior of each known fault event of the generation resource equipment type, trains the stage development recognition model of each fault event, and generates the hyperplane boundary of the development stage of each known fault event of the generation resource equipment type. The fault event triggering risk probability module acquires real-time operating data of several distributed generation resource equipment types in the active distribution network, fits and divides the hyperplane boundary of the known fault event development stage of each generation resource equipment type, determines the real-time operating data of the generation resource equipment type pointing to the known fault event development stage, constructs a real-time fault event triggering risk assessment model, and generates the real-time operating data of several distributed generation resource equipment types in the active distribution network pointing to the known fault event triggering risk probability. The optimal suppression action module, based on the known suppression action database for each fault event of the power generation resource equipment type and the real-time operation data pointing to the known trigger risk probability of each fault event, constructs a deep reinforcement learning controller, establishes an intervention function for the suppression action of the fault event trigger risk probability, and generates the optimal suppression action pointing to the known trigger risk probability of each fault event based on the real-time operation data.

[0018] The hyperplane boundary module specifically includes: The original time-series dataset establishment unit is based on the historical database of the active distribution network. It obtains historical operation data of the active distribution network according to several distributed generation resource equipment types, including historical SCADA data, fault recording data, equipment logs and operation and maintenance reports. It aligns timestamps and performs data preprocessing to establish the original time-series dataset of the historical operation of the active distribution network. The types of distributed power generation equipment include: photovoltaic inverters, wind power converters, and energy storage PCS. Historical SCADA data includes: time-series data of voltage, current, temperature and power in the active distribution network; The fault event set establishment unit extracts the time series of historical fault cases of the active distribution network based on the original time series dataset of the historical operation of the active distribution network. Combining expert experience manuals and operation and maintenance records, it automatically parses the text using named entity recognition technology to generate standardized fault descriptions of historical fault cases of the active distribution network and constructs a historical fault event set of the active distribution network. The causal path generation unit, for each fault event in the historical fault event set of the active distribution network, based on the physical models of several distributed generation resource equipment types and historical fault cases of the active distribution network combined with the expert group, takes each known non-occurring fault event of the generation resource equipment type as the top event of the fault tree, analyzes the known fault factors of the generation resource equipment type that affect the top event of the fault tree and calculates the probability of occurrence of the factors, establishes the fault tree of the fault event of the generation resource equipment type, and obtains the causal path of the development of each known fault event of the generation resource equipment type.

[0019] The hyperplane boundary module also includes: The stage label unit is designed with a fixed sliding time window. Anchor point mapping is used to define the causal path of each known fault event of power generation resource equipment type in the original time series dataset of the active distribution network in the historical operation within the time window and its development performance. The data segments corresponding to each development stage are located for each fault event. Through threshold rules, the data segments are labeled with the corresponding stage labels, including normal stage, budding stage, development stage and adjacent stage. A dataset of known fault events of power generation resource equipment type based on stage labels is established. To further explain, the normal stage includes: no abnormality of any fault factor is detected; the nascent stage includes: more than one fault factor shows a slight abnormality for the first time; the development stage includes: the number of fault factors increases and the degree of abnormality intensifies, and the system state continues to deteriorate; the near stage includes: the abnormal indicators of fault factors reach the fault threshold.

[0020] The hyperplane boundary module also includes: The time-domain feature extraction unit calculates the mean, standard deviation, skewness, kurtosis, root mean square, waveform factor, and peak factor for each data segment in the known fault event dataset based on stage labels of power generation resource equipment types within the time window, obtains the statistical features of each data segment, and extracts its time-domain features. The frequency domain feature extraction unit uses Fast Fourier Transform to convert the time domain signal of each data segment in the known fault event dataset of power generation resource equipment type based on stage label within the time window into a frequency domain signal, and extracts the frequency band energy, dominant frequency component and harmonic distortion rate of each data segment to obtain its frequency domain features. The comprehensive feature vector establishment unit calculates the average slope of the known fault event datasets of power generation resource equipment types based on stage labels within the time window, captures their trend features, combines the time domain features and frequency domain features of each data segment, performs normalization processing, and concatenates them to obtain the comprehensive feature vector of each data segment in the known fault event datasets of power generation resource equipment types based on stage labels.

[0021] The hyperplane boundary module also includes: The dimensionality reduction unit uses principal component analysis to centralize the comprehensive feature vector of each data segment in the known fault event dataset of power generation resource equipment type based on stage labels. It calculates the feature covariance matrix of each data segment, performs eigenvalue decomposition, obtains the feature value and feature vector of each data segment, and selects the top k feature values ​​as principal components to obtain the dimensionality-reduced comprehensive feature vector of each data segment in the known fault event dataset of power generation resource equipment type based on stage labels. The development identification model training unit uses the SVM support vector machine algorithm. It takes the comprehensive feature vector of each data segment in the known fault event dataset of power generation resource equipment type based on stage label after dimensionality reduction and the corresponding stage label as input, and introduces the radial basis function as the kernel function to train the stage development identification model of each fault event and generate the hyperplane boundary of the development stage of each known fault event of power generation resource equipment type.

[0022] When using it, combine the content of the above modules: Traditional active power distribution network fault early warning systems rely heavily on single-point threshold alarms and manual experience analysis, making it difficult to systematically identify the gradual evolution of fault development from multi-dimensional time-series data. They also lack quantitative boundary definition for the initiation, development, and imminent stages of faults, leading to delayed warnings or high false alarm rates. The beneficial effect of this approach lies in constructing a quantitative identification model for the multi-stage development of fault events by integrating fault tree causal analysis and time-frequency domain feature extraction. Utilizing sliding windows and hyperplane boundaries, it achieves dynamic boundary division from normal to fault-critical states, improving the accuracy and timeliness of early fault warnings.

[0023] The fault event triggering risk probability module specifically includes: The real-time operation data comprehensive feature vector extraction unit, based on fieldbus, collects real-time operation data of several distributed generation resource equipment types in the active distribution network, performs data preprocessing, and combines the comprehensive feature vector acquisition technology of each data segment in the known fault event dataset of each generation resource equipment type based on stage label to extract the comprehensive feature vector of the real-time operation data of several distributed generation resource equipment types in the active distribution network within the time window.

[0024] The fault event triggering risk probability module also includes: The fault development stage segmentation unit takes the comprehensive feature vector of real-time operation data of several distributed generation resource equipment types in the active distribution network within the time window as input, associates and pairs it with the stage development identification model of each fault event, calculates the functional interval between the comprehensive feature vector of the real-time operation data and the known hyperplane boundary of each fault event development stage of the generation resource equipment type, and determines the fault development stage to which the comprehensive feature vector of the real-time operation data belongs based on the positive or negative sign of the functional interval. The highest confidence level output by each fault event stage development identification model is selected to determine the real-time operation data of the generation resource equipment type pointing to the known development stages of each fault event.

[0025] The fault event triggering risk probability module also includes: The fault event triggering risk probability determination unit, based on the real-time operation data of the determined power generation resource equipment type pointing to the known development stages of each fault event, calculates the geometric distance from the real-time operation data of the power generation resource equipment type to the hyperplane boundary of the corresponding fault stage, performs mapping and normalization processing, calculates the fault event triggering risk probability value of the real-time operation data of the power generation resource equipment type, constructs a real-time fault event triggering risk assessment model, and generates the known fault event triggering risk probability of several distributed power generation resource equipment types in the active distribution network pointing to the real-time operation data of each fault event.

[0026] When using it, combine the content of the above modules: Current technologies for fault risk assessment in active distribution networks primarily rely on offline analysis based on historical statistical data and threshold-based early warning mechanisms. These methods depend on fixed thresholds and static models, making it difficult to adapt to the real-time dynamic changes in the operating status of distributed generation resources. They lack refined identification of the phased development process of faults, failing to distinguish the transitional characteristics of different fault stages, resulting in insufficient early risk warning capabilities. Traditional methods are often based on single data sources or simple feature extraction, failing to fully integrate the comprehensive real-time operating characteristics of multiple equipment types, thus affecting the accuracy and timeliness of risk probability assessment. This step introduces a fault event dataset feature extraction and dynamic identification model based on phase labels, combined with real-time data preprocessing and comprehensive feature vector extraction, to more comprehensively characterize the operating status of multiple distributed generation resources in active distribution networks. Through a fault event phased development identification model and hyperplane boundary function interval calculation, dynamic and refined discrimination of fault development stages is achieved, improving the sensitivity of early risk identification. Based on a probability calculation method using geometric distance mapping and normalization, a real-time fault event triggering risk assessment model is constructed, generating real-time fault triggering risk probabilities at the equipment type level, improving the real-time performance, accuracy, and adaptability to dynamic operating environments of risk assessment.

[0027] The optimal suppression action module specifically includes: The suppression action database establishment unit integrates the specifications of power generation resource equipment manufacturers, power grid dispatching specifications and expert experience, defines the suppression actions that can be executed at each stage of the development of each known fault type, sets the expected multi-level risk control objectives triggered by each known fault event, and establishes a database of suppression actions for each known fault event of power generation resource equipment type. The executable suppression actions include: alarm, power reduction operation, control mode adjustment, switching to backup equipment, and orderly shutdown; the multi-level risk control targets include: risk probability <0.3 in the nascent stage and risk probability <0.6 in the development stage.

[0028] The optimal suppression action module also includes: The deep reinforcement learning controller construction unit uses the comprehensive feature vector of real-time operation data of several distributed generation resource equipment types in the active distribution network within a time window, the development stage of each known fault event pointed to by the real-time operation data of each generation resource equipment type, and the trigger risk probability of each known fault event pointed to by the real-time operation data as the state space. The effective action subset that conforms to the current fault type in the known fault event suppression action database of each generation resource equipment type is used as the action space. The reduction amount of each fault event trigger risk probability, the generation penalty, and the action frequency penalty are used as the composite reward function. The local system of the active distribution network under edge node control is used as the environment. The deep reinforcement learning controller is constructed to calculate the state of the local system of the active distribution network and the fault event trigger risk probability at the next time step, establish the fault event trigger risk probability suppression action intervention function, and generate the optimal suppression action pointed to by the real-time operation data of each known fault event trigger risk probability.

[0029] When using it, combine the content of the above modules: Traditional fault suppression strategies typically rely on fixed rule bases or single threshold judgments, making it difficult to dynamically adapt to the complex operating states and fault evolution processes of distributed generation resources in active distribution networks. Existing methods are mostly based on offline procedures and experience-based preset actions, lacking quantitative assessment and adaptive decision-making capabilities for real-time risk probabilities. This leads to problems such as delayed suppression actions, excessive intervention, or high action frequency, failing to achieve synergistic optimization of fault suppression and operational economy under multi-level risk control objectives. This step constructs a suppression action module based on deep reinforcement learning, realizing dynamic perception and adaptive decision-making for fault development stages. This module uses real-time operating data feature vectors and fault risk probabilities as the state space, combined with a multi-objective composite reward function, to generate online the optimal suppression strategy that balances risk control, generation loss, and action frequency. This effectively improves the accuracy and economy of fault suppression, while enhancing the active defense capability of active distribution networks against multiple types of fault events.

[0030] Based on the above, the specific implementation method is as follows: In a demonstration area of ​​an active power distribution network, the system was deployed at edge computing nodes, covering more than 200 distributed generation devices in the area, including photovoltaic inverters, wind power converters and energy storage PCS. The system accesses historical SCADA data, fault waveform records, and equipment operation and maintenance logs. Through timestamp alignment and cleaning, a three-year historical time-series dataset was constructed. The expert group combined the equipment physical model with historical fault cases to draw fault trees for typical events such as IGBT overheating faults in photovoltaic inverters and grid-side voltage surges in wind power converters. The causal development path from a slight decrease in cooling fan speed to a core temperature exceeding the critical value was clarified. Using sliding time windows and threshold rules, historical data segments were labeled in four stages, including normal, nascent, developing, and imminent, forming a standardized fault event dataset for model training. During operation, the system collects voltage, current, power, and temperature data of each device in real time via fieldbus. The hyperplane boundary module extracts time-frequency domain features from historical data and performs principal component dimensionality reduction. It uses SVM support vector machine to train and generate stage identification models for each fault event, outputting a clear hyperplane decision boundary. The system calculates the comprehensive feature vector of real-time data every 5 seconds and inputs it into the relevant fault identification model. By calculating the functional interval between the feature vector and the hyperplane boundary, it determines the current fault development stage of the device. It identifies that the photovoltaic inverter is in the development stage of IGBT overheating fault. By calculating the geometric distance from the feature vector to the hyperplane boundary of this stage, mapping and normalizing, it outputs the real-time trigger risk probability value of the fault event as 0.72. When the risk probability value exceeds the development stage target value of 0.6, the optimal suppression action module is immediately activated. The deep reinforcement learning controller of this module takes the feature vector of the current state, the fault stage, and the risk probability of 0.72 as inputs, and selects an action from the preset action library, such as alarm, power reduction of 15%, and switching to backup cooling device. The controller takes reducing the risk probability as its core objective, while taking into account minimizing power generation loss and avoiding frequent action switching. Through interactive learning with the environment as a local system of the distribution network, it outputs the current optimal suppression action, executes a power reduction of 10% and starts enhanced cooling. After the system executes this action, it continuously monitors the change in risk probability, forming a closed loop of perception-evaluation-decision-execution. Actual operation shows that the system advances the warning time of critical faults by an average of 40 minutes, reduces the false alarm rate by about 35%, and successfully avoids multiple unplanned outages in the early stage of faults through precise suppression actions, thereby improving the reliability and economy of power distribution network operation.

[0031] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. An active power distribution network anomaly sensing system based on edge computing, characterized in that, include: Hyperplane boundary module, fault event triggering risk probability module, optimal suppression action module; The fault event triggering risk probability module is electrically connected to the hyperplane boundary module, and the optimal suppression action module is electrically connected to the fault event triggering risk probability module. The hyperplane boundary module acquires prior knowledge of the known triggering behavior of each fault event of several distributed generation resource equipment types in the active distribution network, constructs the causal path of the development behavior of each known fault event of the generation resource equipment type, trains the stage development recognition model of each fault event, and generates the hyperplane boundary of the development stage of each known fault event of the generation resource equipment type. The fault event triggering risk probability module acquires real-time operating data of several distributed generation resource equipment types in the active distribution network, fits and divides the hyperplane boundary of the known fault event development stage of each generation resource equipment type, determines the real-time operating data of the generation resource equipment type pointing to the known fault event development stage, constructs a real-time fault event triggering risk assessment model, and generates the real-time operating data of several distributed generation resource equipment types in the active distribution network pointing to the known fault event triggering risk probability. The optimal suppression action module, based on the known suppression action database for each fault event of the power generation resource equipment type and the real-time operation data pointing to the known trigger risk probability of each fault event, constructs a deep reinforcement learning controller, establishes an intervention function for the suppression action of the fault event trigger risk probability, and generates the optimal suppression action pointing to the known trigger risk probability of each fault event based on the real-time operation data.

2. The active distribution network anomaly sensing system based on edge computing according to claim 1, characterized in that, The hyperplane boundary module specifically includes: The original time-series dataset establishment unit is based on the historical database of the active distribution network. It obtains historical operation data of the active distribution network according to several distributed generation resource equipment types, including historical SCADA data, fault recording data, equipment logs and operation and maintenance reports. It aligns timestamps and performs data preprocessing to establish the original time-series dataset of the historical operation of the active distribution network. The types of distributed power generation equipment include: photovoltaic inverters, wind power converters, and energy storage PCS. Historical SCADA data includes: time-series data of voltage, current, temperature and power in the active distribution network; The fault event set establishment unit extracts the time series of historical fault cases of the active distribution network based on the original time series dataset of the historical operation of the active distribution network. Combining expert experience manuals and operation and maintenance records, it automatically parses the text using named entity recognition technology to generate standardized fault descriptions of historical fault cases of the active distribution network and constructs a historical fault event set of the active distribution network. The causal path generation unit, for each fault event in the historical fault event set of the active distribution network, based on the physical models of several distributed generation resource equipment types and historical fault cases of the active distribution network combined with the expert group, takes each known non-occurring fault event of the generation resource equipment type as the top event of the fault tree, analyzes the known fault factors of the generation resource equipment type that affect the top event of the fault tree and calculates the probability of occurrence of the factors, establishes the fault tree of the fault event of the generation resource equipment type, and obtains the causal path of the development of each known fault event of the generation resource equipment type.

3. The active distribution network anomaly sensing system based on edge computing according to claim 2, characterized in that, The hyperplane boundary module also includes: The stage label unit is designed with a fixed sliding time window. Anchor point mapping is used to define the causal path of each known fault event of power generation resource equipment type in the original time series dataset of the active distribution network in the historical operation within the time window and its development performance. The data segments corresponding to each development stage are located for each fault event. Through threshold rules, the data segments are labeled with the corresponding stage labels, including normal stage, budding stage, development stage and adjacent stage. A dataset of known fault events of power generation resource equipment type based on stage labels is established. To further explain, the normal stage includes: no abnormality of any fault factor is detected; the nascent stage includes: more than one fault factor shows a slight abnormality for the first time; the development stage includes: the number of fault factors increases and the degree of abnormality intensifies, and the system state continues to deteriorate; the near stage includes: the abnormal indicators of fault factors reach the fault threshold.

4. The active distribution network anomaly sensing system based on edge computing according to claim 3, characterized in that, The hyperplane boundary module also includes: The time-domain feature extraction unit calculates the mean, standard deviation, skewness, kurtosis, root mean square, waveform factor, and peak factor for each data segment in the known fault event dataset based on stage labels of power generation resource equipment types within the time window, obtains the statistical features of each data segment, and extracts its time-domain features. The frequency domain feature extraction unit uses Fast Fourier Transform to convert the time domain signal of each data segment in the known fault event dataset of power generation resource equipment type based on stage label within the time window into a frequency domain signal, and extracts the frequency band energy, dominant frequency component and harmonic distortion rate of each data segment to obtain its frequency domain features. The comprehensive feature vector establishment unit calculates the average slope of the known fault event datasets of power generation resource equipment types based on stage labels within the time window, captures their trend features, combines the time domain features and frequency domain features of each data segment, performs normalization processing, and concatenates them to obtain the comprehensive feature vector of each data segment in the known fault event datasets of power generation resource equipment types based on stage labels.

5. The active distribution network anomaly sensing system based on edge computing according to claim 4, characterized in that, The hyperplane boundary module also includes: The dimensionality reduction unit uses principal component analysis to centralize the comprehensive feature vector of each data segment in the known fault event dataset of power generation resource equipment type based on stage labels. It calculates the feature covariance matrix of each data segment, performs eigenvalue decomposition, obtains the feature value and feature vector of each data segment, and selects the top k feature values ​​as principal components to obtain the dimensionality-reduced comprehensive feature vector of each data segment in the known fault event dataset of power generation resource equipment type based on stage labels. The development identification model training unit uses the SVM support vector machine algorithm. It takes the comprehensive feature vector of each data segment in the known fault event dataset of power generation resource equipment type based on stage label after dimensionality reduction and the corresponding stage label as input, and introduces the radial basis function as the kernel function to train the stage development identification model of each fault event and generate the hyperplane boundary of the development stage of each known fault event of power generation resource equipment type.

6. The active distribution network anomaly sensing system based on edge computing according to claim 1, characterized in that, The fault event triggering risk probability module specifically includes: The real-time operation data comprehensive feature vector extraction unit, based on fieldbus, collects real-time operation data of several distributed generation resource equipment types in the active distribution network, performs data preprocessing, and combines the comprehensive feature vector acquisition technology of each data segment in the known fault event dataset of each generation resource equipment type based on stage label to extract the comprehensive feature vector of the real-time operation data of several distributed generation resource equipment types in the active distribution network within the time window.

7. The active distribution network anomaly sensing system based on edge computing according to claim 6, characterized in that, The fault event triggering risk probability module also includes: The fault development stage segmentation unit takes the comprehensive feature vector of real-time operation data of several distributed generation resource equipment types in the active distribution network within the time window as input, associates and pairs it with the stage development identification model of each fault event, calculates the functional interval between the comprehensive feature vector of the real-time operation data and the known hyperplane boundary of each fault event development stage of the generation resource equipment type, and determines the fault development stage to which the comprehensive feature vector of the real-time operation data belongs based on the positive or negative sign of the functional interval. The highest confidence level output by each fault event stage development identification model is selected to determine the real-time operation data of the generation resource equipment type pointing to the known development stages of each fault event.

8. The active distribution network anomaly sensing system based on edge computing according to claim 7, characterized in that, The fault event triggering risk probability module also includes: The fault event triggering risk probability determination unit, based on the real-time operation data of the determined power generation resource equipment type pointing to the known development stages of each fault event, calculates the geometric distance from the real-time operation data of the power generation resource equipment type to the hyperplane boundary of the corresponding fault stage, performs mapping and normalization processing, calculates the fault event triggering risk probability value of the real-time operation data of the power generation resource equipment type, constructs a real-time fault event triggering risk assessment model, and generates the known fault event triggering risk probability of several distributed power generation resource equipment types in the active distribution network pointing to the real-time operation data of each fault event.

9. The active distribution network anomaly sensing system based on edge computing according to claim 8, characterized in that, The optimal suppression action module specifically includes: The suppression action database establishment unit integrates the specifications of power generation resource equipment manufacturers, power grid dispatching specifications and expert experience, defines the suppression actions that can be executed at each stage of the development of each known fault type, sets the expected multi-level risk control objectives triggered by each known fault event, and establishes a database of suppression actions for each known fault event of power generation resource equipment type. The executable suppression actions include: alarm, power reduction operation, control mode adjustment, switching to backup equipment, and orderly shutdown; the multi-level risk control targets include: risk probability <0.3 in the nascent stage and risk probability <0.6 in the development stage.

10. The active distribution network anomaly sensing system based on edge computing according to claim 9, characterized in that, The optimal suppression action module also includes: The deep reinforcement learning controller construction unit uses the comprehensive feature vector of real-time operation data of several distributed generation resource equipment types in the active distribution network within a time window, the development stage of each known fault event pointed to by the real-time operation data of each generation resource equipment type, and the trigger risk probability of each known fault event pointed to by the real-time operation data as the state space. The effective action subset that conforms to the current fault type in the known fault event suppression action database of each generation resource equipment type is used as the action space. The reduction amount of each fault event trigger risk probability, the generation penalty, and the action frequency penalty are used as the composite reward function. The local system of the active distribution network under edge node control is used as the environment. The deep reinforcement learning controller is constructed to calculate the state of the local system of the active distribution network and the fault event trigger risk probability at the next time step, establish the fault event trigger risk probability suppression action intervention function, and generate the optimal suppression action pointed to by the real-time operation data of each known fault event trigger risk probability.

Images