Event-driven architecture-oriented spatiotemporal feature extraction method and system for multi-type fault power outage events

Abstract

The application discloses a method and system for extracting space-time characteristics of multi-type fault outage events for an event-driven architecture, and belongs to the field of power system fault diagnosis. The method collects historical fault data, equipment operation data and external environment data through event monitoring triggering and multi-source data fusion. An XGBoost model based on probability distribution is used to accurately clean and complete the data, and a comprehensive all-scenario fault sample library is constructed. Then, the time series convolution network and the graph convolution neural network are used to deeply mine and fuse the space-time characteristics of the fault events, such as periodicity, high-incidence period, geographical distribution rule and spatial correlation. The application can efficiently and intelligently realize the analysis of the space-time evolution rule of the fault events and the identification of the key influence factors, significantly improve the accuracy and timeliness of the power grid fault research and judgment, provide a scientific basis for fault positioning and emergency disposal, and help the construction of a new type of power system.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power system fault diagnosis, in particular to a spatio-temporal feature extraction method and system for multi-type fault outage events based on an event-driven architecture. BACKGROUND

[0002] As a key link of the power system directly facing terminal users, the safe and stable operation of the distribution network is of great importance. With the expansion of the power grid and the increasing demand for power supply reliability from users, it has become a core requirement for intelligent operation and maintenance of the power grid to quickly and accurately analyze fault outage events of the distribution network. The analysis of outage events aims to extract fault features, identify fault types, and mine their occurrence rules by analyzing massive operation data, thereby providing decision support for fault handling and preventive maintenance, which usually involves processing and analysis techniques for large-scale, multi-dimensional data.

[0003] In the prior art, the analysis method for outage events of the distribution network usually relies on a single data processing model or analysis dimension. Some methods focus on using time series analysis models such as autoregressive integrated moving average models or recurrent neural networks to predict the occurrence probability of outage events or identify their periodicity. Other methods use spatial statistics or clustering algorithms to analyze the distribution characteristics of outage events in geographical locations to identify high-fault areas. In the data preprocessing stage, simple statistical methods or interpolation methods are usually used to handle noise and missing values in the data.

[0004] However, the existing technical solutions have obvious deficiencies in practical application. First, the time characteristics and spatial characteristics are analyzed separately, ignoring the complex dynamic coupling relationship between them, making it difficult to truly reflect the spatio-temporal evolution process of faults in the power grid. Second, the existing models have poor adaptability to multi-source, heterogeneous, and high-noise data of the distribution network, and the data cleaning effect is not good, which limits the accuracy of feature extraction. In addition, most methods use offline batch processing mode, which lacks real-time response capability to sudden fault events, making it difficult to meet the demand of modern power grids for rapid fault handling. SUMMARY

[0005] To solve the above problems, the present application provides a spatio-temporal feature extraction method and system for multi-type fault outage events based on an event-driven architecture, which uses an XGBoost model based on probability distribution to clean multi-source data cooperatively, and deeply integrates time series convolutional networks and graph convolutional neural networks to extract spatio-temporal features, thereby achieving real-time and accurate analysis and rule mining of fault outage events.

[0006] The above objectives can be achieved by the following solutions:

[0007] The method for extracting spatiotemporal features of multi-type fault outage events of an event-driven architecture comprises: listening to fault outage event signals in a power distribution network; when an event occurs, triggering a data collection process; the data collection process is to collect multi-source fault outage event data of historical fault outage event data, equipment operation data and external environment data of the power distribution network to obtain an original data set; performing XGBoost model cleaning based on probability distribution on the original data set to identify and remove noise and outliers, complete missing data, and obtain a standard data set; based on the standard data set, constructing a power distribution network full-scenario fault outage sample library; based on the fault outage sample library, using a time series convolution network algorithm to extract time features of the fault outage event data, modeling and analyzing periodicity and high-occurrence periods of the fault outage event to obtain a time feature vector; based on the fault outage sample library, using a graph convolution neural network algorithm to extract spatial features of the fault outage event data, revealing geographical distribution rules and spatial correlations of different regions, equipment and lines to obtain a spatial feature vector; dynamically weighting and fusing the time feature vector and the spatial feature vector to obtain a comprehensive spatiotemporal feature vector; based on the comprehensive spatiotemporal feature vector, identifying spatiotemporal evolution rules and key influencing factors of the fault outage event.

[0008] Optionally, the method further comprises: setting an event listener to monitor event logs and alarm signals in a power distribution network monitoring system to obtain an event trigger signal; verifying the event trigger signal according to a verification rule for distinguishing real faults from false alarms to obtain a valid event; initializing data collection parameters based on the valid event to trigger the data collection process.

[0009] Optionally, the method further comprises: optimizing the fault outage sample library, wherein: using the evolution rules and key influencing factors to feed back and update the power distribution network full-scenario fault outage sample library; based on the updated fault outage sample library, retraining the time series convolution network and the graph convolution neural network to optimize feature extraction performance; iteratively performing the feedback updating and retraining steps to realize collaborative optimization of the fault outage sample library and the model.

[0010] Optionally, the XGBoost model cleaning based on probability distribution on the original data set to identify and remove noise and outliers and complete missing data comprises: using an XGBoost model to calculate importance scores of each feature in the original data set to obtain a feature importance list; according to the feature importance list, screening features with importance scores satisfying a predetermined screening condition to form a key feature subset; performing probability distribution estimation on the key feature subset to generate a joint probability distribution model; using the joint probability distribution model to detect and remove noise and outliers in the original data set while completing missing data, and outputting a standard data set.

[0011] Optionally, based on the fault power failure sample library, time feature extraction of the fault power failure event data by using a time sequence convolution network algorithm comprises: obtaining time sequence data from the fault power failure sample library, including power failure event sequences in day, week and season dimensions, to obtain time sequence input; performing causal convolution and dilated convolution processing on the time sequence input by using a time sequence convolution network, to capture long-term dependence, to obtain preliminary time features; performing threshold mechanism adjustment and residual connection optimization on the preliminary time features, to enhance feature stability, to output a time feature vector.

[0012] Optionally, based on the fault power failure sample library, spatial feature extraction of the fault power failure event data by using a graph convolution neural network algorithm comprises: obtaining power distribution network topology structure data from the fault power failure sample library, including connection relationships of stations, lines, transformers and users, to obtain graph structure input; performing neighbor node feature aggregation on the graph structure input by using a graph convolution neural network, to learn spatial correlation, to obtain preliminary spatial features; performing normalization processing and layer-by-layer propagation on the preliminary spatial features, to expand the receptive field, to output a spatial feature vector.

[0013] Optionally, dynamic weighted fusion of the time feature vector and the spatial feature vector comprises:

[0014] obtaining a dynamic weight value, the dynamic weight value being used to reflect the influence degree of the time feature vector and the spatial feature vector on the fault power failure event; performing weighted summation on the time feature vector and the spatial feature vector by using the dynamic weight value, to obtain fused features; performing dimension reduction and standardization processing on the fused features, to output a comprehensive spatio-temporal feature vector.

[0015] Optionally, based on the comprehensive spatio-temporal feature vector, identifying the spatio-temporal evolution law of the fault power failure event and its key influencing factors comprises: obtaining causes, influence degrees and recovery characteristics of the fault power failure event; performing law mining on the causes, influence degrees and recovery characteristics and the comprehensive spatio-temporal feature vector; based on the law mining result, revealing the evolution law of different types of fault power failure events in time and space, and identifying key factors influencing the event.

[0016] Based on the same inventive concept, the application also provides a space-time feature extraction system for multi-type fault power outage events, comprising: an event monitoring and collecting module, configured to monitor fault power outage event signals in a power distribution network, and when an event occurs, trigger a collection process to collect multi-source fault power outage event data of historical fault power outage event data, device operation data and external environment data of the power distribution network, to obtain an original data set; a data cleaning module, configured to clean the original data set based on an XGBoost model based on probability distribution, identify and remove noise and outliers, and complete missing data, to obtain a standard data set; a sample library construction module, configured to construct a power distribution network full-scene fault power outage sample library based on the standard data set; a time feature extraction module, configured to extract time features of the fault power outage event data based on the fault power outage sample library using a time series convolution network algorithm, model and analyze the periodicity and high-occurrence period of the fault power outage event, to obtain a time feature vector; a spatial feature extraction module, configured to extract spatial features of the fault power outage event data based on the fault power outage sample library using a graph convolution neural network algorithm, to reveal the geographical distribution law and spatial correlation of different regions, devices and lines, to obtain a spatial feature vector; a feature fusion module, configured to dynamically weight and fuse the time feature vector and the spatial feature vector, to obtain a comprehensive space-time feature vector; and a law mining module, configured to identify the space-time evolution law of the fault power outage event and its key influencing factors based on the comprehensive space-time feature vector.

[0017] Compared with the prior art, the application has the following advantages:

[0018] 1. The application realizes real-time response and intelligent analysis of power distribution network fault power outage events by constructing an event-driven data processing architecture. When a fault event occurs, the parallel collection of multi-source data can be triggered immediately, and a series of automatic data cleaning, feature extraction and law mining processes can be performed to provide immediate support for rapid fault diagnosis and decision-making, significantly improving the timeliness and initiative of fault handling and effectively shortening the power outage response time.

[0019] 2. The application combines time series convolution networks and graph convolution neural networks to extract the space-time features of power outage events in depth. The time series model is responsible for mining the time evolution law of the fault, and the graph model is responsible for revealing the spatial conduction relationship of the fault in the power grid topology. The two models work together to capture complex coupling patterns that are ignored by traditional single-dimensional analysis methods, greatly improving the accuracy and depth of fault analysis, and making the prediction of fault causes, impact range and development trend more reliable.

[0020] 3、The application has the ability of self-evolution and continuous learning through the design of a closed-loop optimization process containing feedback update and model retraining, which can continuously learn new knowledge from new fault events and use these knowledge to optimize the data sample library and analysis model, thereby ensuring that the method can continuously adapt to changes in power grid structure and the emergence of new fault modes, maintaining the long-term effectiveness and advancement of the technical solution.

[0021] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art upon examination of the following or can be learned by practice of the present application. The objects and other advantages of the present application can be realized and attained by the structure particularly pointed out in the description, claims and drawings. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application or in the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0023] Figure 1 is a flowchart of the event-driven architecture-oriented spatiotemporal feature extraction method of the multi-type fault outage event of the embodiment of the present application.

[0024] Figure 2 is an annual time sequence distribution graph of the fault outage event of the embodiment of the present application.

[0025] Figure 3 is a spatial distribution heat map of the fault outage event of the embodiment of the present application.

[0026] Figure 4 is a structural schematic diagram of a spatiotemporal feature extraction system of a multi-type fault outage event of the embodiment of the present application. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be clearly and completely described below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0028] REFERENCE Figure 1An embodiment of the present application provides a method for extracting spatio-temporal features of multi-type fault outage events in an event-driven architecture, adopts an XGBoost model based on probability distribution to clean multi-source data, and deeply fuses a time series convolution network and a graph convolution neural network to extract spatio-temporal features, so that real-time and accurate judgment and law mining of the fault outage events can be realized, and timeliness, accuracy and initiative of fault handling are significantly improved.

[0029] The method of the embodiment specifically comprises:

[0030] A fault outage event signal in a power distribution network is listened to, and when the event occurs, a data collection process is triggered;

[0031] The data collection process is to collect multi-source fault outage event data of historical fault outage event data, equipment operation data and external environment data of the power distribution network, to obtain an original data set;

[0032] The original data set is cleaned by an XGBoost model based on probability distribution, noise and outliers are identified and removed, missing data is completed, and a standard data set is obtained;

[0033] Based on the standard data set, a power distribution network full-scene fault outage sample library is constructed;

[0034] Based on the fault outage sample library, a time series convolution network algorithm is used to extract time features of the fault outage event data, to model and analyze periodicity and high-incidence periods of the fault outage events, and to obtain a time feature vector;

[0035] Based on the fault outage sample library, a graph convolution neural network algorithm is used to extract space features of the fault outage event data, to reveal geographical distribution rules and spatial correlation of different regions, equipment and lines, and to obtain a space feature vector;

[0036] The time feature vector and the space feature vector are dynamically weighted and fused to obtain a comprehensive spatio-temporal feature vector;

[0037] Based on the comprehensive spatio-temporal feature vector, spatio-temporal evolution rules and key influence factors of the fault outage events are identified.

[0038] Specifically, by listening to the fault outage event signal in the power distribution network, the data collection process is triggered. The historical fault outage event data, equipment operation data and external environment data are collected to obtain a multi-source original data set. Based on the original data set, the XGBoost model is used for accurate cleaning and completion to build a comprehensive power distribution network full-scenario fault outage sample library. This sample library is dynamic, which combines static fault history data with real-time changing equipment and environmental parameters to provide high-quality benchmarks for subsequent analysis. Based on the sample library, the time series convolution network and the graph convolution neural network are used to deeply mine and dynamically weight the periodicity of the fault outage event and the high-occurrence period and the geographical distribution law of different regions, equipment and lines. The seemingly independent spatio-temporal information is effectively integrated to generate a comprehensive spatio-temporal feature vector that can fully represent the fault event. Based on the comprehensive spatio-temporal feature vector, the spatio-temporal evolution law of the fault outage event and its key influencing factors are identified. These spatio-temporal features are used as decision-making basis to reveal the evolution patterns of different types of fault events in time and space, thereby realizing continuous, dynamic and adaptive prediction and analysis of faults and improving the operation efficiency of the power grid.

[0039] Optionally, the listening to the fault outage event signal in the power distribution network comprises:

[0040] An event listener is set to monitor the event log and alarm signal in the power distribution network monitoring system to obtain an event trigger signal.

[0041] Specifically, by setting an event listener, the event log and alarm signal from the power distribution network monitoring system are monitored in real time to obtain initial signals reflecting abnormal conditions of the power grid, including but not limited to device failure, voltage out-of-limit or current surge, etc., thereby obtaining an event trigger signal.

[0042] Illustratively, an event listener is set to monitor the log sent by the power distribution network monitoring center in real time. The log shows that a short circuit occurs in a line in a certain area, i.e., an event trigger signal is obtained.

[0043] The event trigger signal is verified according to a verification rule for distinguishing real faults from false alarms to obtain an effective event.

[0044] Specifically, the event trigger signal is verified according to a preset verification rule to exclude false alarms and transient interference. The verification rule can be a logic based on threshold, duration or multi-signal cooperative judgment. If the event trigger signal meets the verification rule, the event is determined to be a real fault, and an effective event is obtained.

[0045] Illustratively, suppose the verification rule is set as: when the voltage of a line is continuously lower than the safety threshold and exceeds 500 milliseconds, it is considered as an effective event that needs to be handled.

[0046] Based on the effective event, initialize data collection parameters, trigger data collection process.

[0047] Specifically, based on the fault information contained in the effective event, initialize data collection parameters. These parameters include the time, location, type of fault occurrence, and affected device number, etc. Then, use these parameters to trigger the data collection process and start the collection of raw data.

[0048] Exemplarily, based on the effective event, the fault line number, fault occurrence time and fault type are taken as data collection parameters, and the data collection process is initialized to collect relevant data of the line before and after the fault.

[0049] Optionally, the method further comprises: optimization of the fault outage sample library, wherein:

[0050] Using the evolution law and key influencing factors, feedback updates the power distribution network full-scenario fault outage sample library;

[0051] Specifically, using the evolution law and key influencing factors obtained from the analysis, these information is used as feedback to update the labels and metadata of corresponding fault events in the power distribution network full-scenario fault outage sample library. This ensures the continuous accuracy and effectiveness of the sample library.

[0052] Exemplarily, when the algorithm identifies that the frequency of a certain type of line fault significantly increases under extreme high temperature, "high temperature weather" is added as a key influencing factor to the metadata of all related historical fault events, and the labels are updated accordingly.

[0053] Based on the updated fault outage sample library, retrain the time convolution network and graph convolution neural network to optimize the feature extraction performance;

[0054] Specifically, based on the updated fault outage sample library, the time convolution network and graph convolution neural network are retrained to optimize their feature extraction performance. This process enables the model to adapt to new fault patterns and data features, thereby improving the overall research and judgment ability.

[0055] Exemplarily, using the newly fed 50 fault sample data containing the latest extreme weather influence, the time convolution network and graph convolution neural network are retrained, so that the prediction accuracy of the model for similar fault scenarios is improved by about 5%.

[0056] Iteratively perform the feedback update and retraining steps to achieve the coordinated optimization of the fault outage sample library and the model.

[0057] Specifically, by iteratively performing the above feedback updating and retraining steps, the synergy of the fault outage sample library and the model is achieved. This process forms a closed-loop adaptive mechanism, enabling data and models to learn from each other, co-evolve, and ensure the long-term effectiveness and advancement of the technical solution.

[0058] Illustratively, the sample library update and model retraining iteration is automatically performed every other week or when the newly collected fault data accumulates more than 100, ensuring that the model always remains adaptable to the latest state of the power grid.

[0059] Optionally, the original data set is subjected to XGBoost model cleaning based on probability distribution, identifying and removing noise and outliers, and filling in missing data, including:

[0060] The importance score of each feature in the original data set is calculated using the XGBoost model, resulting in a feature importance list.

[0061] Specifically, the built-in feature importance analysis capability of the XGBoost model is used to train and evaluate all features in the original data set, calculate the contribution of each feature to fault analysis, and obtain a feature importance list reflecting the importance of each feature.

[0062] Illustratively, the XGBoost model analysis results show that the importance scores of voltage, current mutation amplitude, and meteorological conditions in the region where the fault point is located are much higher than historical maintenance records and device models, so the top three are considered high-importance features.

[0063] According to the feature importance list, features with importance scores meeting predetermined filtering conditions are filtered to form a key feature subset.

[0064] Specifically, according to the feature importance list, a predetermined filtering condition is set, such as an importance score greater than 0.80, to filter out the features that contribute most to fault analysis, thereby forming a compact and efficient key feature subset that removes redundant and irrelevant features.

[0065] Illustratively, features such as fault current, fault time, and fault type with importance scores higher than 0.80 are filtered out from the original data set to form a key feature subset containing core information.

[0066] Probability distribution estimation is performed on the key feature subset to generate a joint probability distribution model.

[0067] Specifically, probability distribution estimation is performed on the filtered key feature subset to capture the relationships and distribution patterns between these features, thereby generating a joint probability distribution model that describes the normal data pattern.

[0068] Exemplarily, a multivariate Gaussian probability distribution estimation is performed on a key feature subset including voltage, current and temperature, a model describing the joint distribution of these variables in normal operation state is established, which can be used for subsequent anomaly data detection.

[0069] Using the joint probability distribution model, noise and outliers in the original data set are detected and removed, and missing data is completed, and a standard data set is output.

[0070] Specifically, using the joint probability distribution model, by calculating the probability density of each data point under the model, noise and outliers with extremely low probability are detected and removed. At the same time, for the missing data in the data set, the joint probability distribution model is used for interpolation, and finally a high-quality standard data set is obtained.

[0071] Exemplarily, using the joint probability distribution model, it is found that the probability of some voltage readings compared with normal data patterns is less than 0.001, so these readings are judged as outliers and removed. At the same time, for the missing temperature data, it is calculated according to the joint distribution with voltage and current to complete the data.

[0072] Optionally, based on the fault power failure sample library, a time feature extraction is performed on the fault power failure event data using a time convolution network algorithm, comprising:

[0073] Obtaining time series data from the fault power failure sample library, including daily, weekly and seasonal power failure event sequences, obtaining time series input;

[0074] Specifically, time series data is obtained from the fault power failure sample library, which includes power failure event sequences with day, week and season as dimensions. It fully reflects the occurrence frequency, duration and high incidence period of events at different time scales, so as to obtain time series input for time series modeling, which can fully characterize the dynamic characteristics of fault events changing with time.

[0075] Exemplarily, the number of power failures per day or the total number of high-temperature power failure events per month in a certain area in the past two years is obtained as time series input.

[0076] Using a time convolution network to perform causal convolution and dilated convolution on the time series input to capture long-term dependencies, obtaining preliminary time features;

[0077] Specifically, the time series input is processed by causal convolution and dilated convolution using a time convolution network. Causal convolution ensures that the model output does not depend on future data, maintaining the time sequence of events. Dilated convolution can effectively expand the receptive field of the model without increasing the number of model parameters and computational complexity by introducing holes in the convolution kernel, thereby capturing long-term dependencies between events. The calculation formula of dilated convolution is:

[0078] ,

[0079] wherein, represents dilated convolution, is a specific time point, is an input vector, is a convolution kernel with a size of , is a dilation factor that controls the expansion size. Through these processes, preliminary time features reflecting the time evolution rule can be obtained. As shown in Figure 2 , the number of daily occurrences of a fault outage event within a year is shown, where seasonal changes and periodic high-occurrence periods are clearly visible. These time series features are the basis for deep analysis by the time convolution network algorithm.

[0080] For example, a convolution kernel with a size of 2 and a dilation factor of 2 are used. When calculating the output at a specific time point , the convolution kernel will act on the data at positions and of the input vector to capture features of events that are far apart.

[0081] The preliminary time features are adjusted by a threshold mechanism and optimized by residual connection to enhance feature stability, and a time feature vector is output.

[0082] Specifically, the preliminary time features are adjusted by a threshold mechanism to enhance the modeling ability of the network, and residual connections are introduced to optimize the network structure. Residual connections help to alleviate the problem of gradient vanishing or gradient explosion in deep network training, thereby enhancing the stability of the features. Through these optimization steps, the output time feature vector can have higher robustness and reliability.

[0083] For example, after the preliminary time features are processed by the threshold mechanism, the processing results are added to the original preliminary time features through residual connection, so that the network can learn the incremental changes while retaining the original feature information, thereby ensuring the stability of the features.

[0084] Optionally, based on the fault power failure sample library, the spatial feature extraction on the fault power failure event data is performed by using a graph convolutional neural network algorithm, and the spatial feature extraction includes:

[0085] Topology structure data of the power distribution network is obtained from the fault power failure sample library, including connection relationships of stations, lines, transformers and users, to obtain a graph structure input;

[0086] Specifically, topology structure data of the power distribution network is obtained from the fault power failure sample library, including connection relationships of stations, lines, transformers and users in the power grid, and multi-source heterogeneous information associated therewith, such as fault types, geographic coordinates and device attributes, and the power grid structure is represented as a graph based on the topology structure data, to obtain a graph structure input for spatial analysis. The graph structure input can comprehensively reflect the spatial connection characteristics of the power grid and multi-dimensional event information.

[0087] For example, connection relationships of a substation and 10 lines connected thereto, and connection relationships of the lines and downstream transformers and users are obtained, to obtain a graph structure input describing the topology of the power grid.

[0088] The graph structure input is subjected to neighbor node feature aggregation by using a graph convolutional neural network, to learn spatial correlation, and to obtain preliminary spatial features;

[0089] Specifically, the graph structure input is subjected to neighbor node feature aggregation by using a graph convolutional neural network, to learn spatial correlation between nodes of the power grid. The core idea is to simplify the graph convolution operation to a first-order approximation, to capture the propagation relationship of the fault in the topology of the power grid by aggregating feature information of each center node and directly connected neighbor nodes. The core operation is represented as:

[0090]

[0091] is a feature vector matrix of a node at a th layer, is an adjacency matrix with a self-loop, is a diagonal degree matrix thereof, is a learnable weight matrix, is an activation function. The operation performs weighted summation on neighbor node features of each node, to obtain preliminary spatial features reflecting local spatial correlation. As shown in FIG. 6, density distribution of the fault power failure event in the geographic space is intuitively displayed in the form of a heat map, wherein a region with a darker color represents a high-fault area. These spatial distribution rules are key to learning spatial correlation by the graph convolutional neural network algorithm. Figure 3

[0092] ​​​​Exemplarily, one transformer node aggregates the fault features, such as the average outage duration and the fault frequency, from adjacent line nodes and user nodes, and takes these aggregated features as its own preliminary spatial feature vector.

[0093] The preliminary spatial features are normalized and stacked and propagated to expand the receptive field, and a spatial feature vector is output.

[0094] Specifically, the preliminary spatial features are normalized and stacked and propagated to avoid numerical instability and gradient problems. By stacking multiple graph convolution layers, neighborhood information can be gradually propagated to more distant nodes, thereby effectively expanding the receptive field, enabling the model to capture spatial correlations in a wider range, and finally outputting a spatial feature vector that can comprehensively reflect the geographical distribution and spatial conduction law of fault events.

[0095] Exemplarily, a two-layer graph convolution network is used, the first layer aggregates the features of directly connected neighbor nodes, and the second layer aggregates the features of two-hop neighbors, thereby expanding the receptive field from a single line to multiple adjacent lines, and thus being able to capture spatial correlations in a wider range.

[0096] Optionally, the dynamic weighting fusion of the time feature vector and the spatial feature vector comprises:

[0097] A dynamic weight value is obtained, which is used to reflect the influence degree of the time feature vector and the spatial feature vector on the fault outage event;

[0098] Specifically, a special attention network is used to automatically learn and obtain the dynamic weight value. This attention network can adaptively assign weights to the time feature vector and the spatial feature vector according to the context information such as the type, occurrence time and location of the fault outage event, accurately reflecting the relative influence degree of each dimension feature on the fault in different scenarios. The weight value can provide a scientific basis for subsequent feature fusion.

[0099] Exemplarily, for a fault outage event caused by extreme weather such as a typhoon, since its time evolution law is more significant, such as high incidence in a specific season and period, the attention network will assign a higher weight to the time feature vector, for example, assigning a weight of 0.7 to the time feature vector, and the weight of the spatial feature vector is 0.3.

[0100] The time feature vector and the spatial feature vector are weighted and summed using the dynamic weight value to obtain a fusion feature;

[0101] Specifically, the time feature vector and the spatial feature vector are weighted and summed using the obtained dynamic weight value, thereby obtaining a fusion feature that integrates temporal and spatial information. The mathematical expression of this process is represented as:

[0102] ,

[0103] wherein, is the fusion feature, is the temporal feature vector, is the spatial feature vector, and are the corresponding dynamic weight values, respectively, and .

[0104] Exemplarily, assuming the temporal feature vector is , the spatial feature vector is , and the dynamic weight values are and , respectively. Then the fusion feature will be calculated as: .

[0105] Dimension reduction and standardization are performed on the fusion feature to output a comprehensive spatio-temporal feature vector.

[0106] Specifically, dimension reduction is performed on the fusion feature to eliminate redundant information in the data and reduce the computational complexity of the subsequent model. Subsequently, the dimension-reduced feature is standardized, such as Z-Score standardization, to ensure that all features have the same scale, thereby eliminating the influence of different feature units and dimensions on the model performance, and finally outputting a high-quality comprehensive spatio-temporal feature vector.

[0107] Exemplarily, a 200-dimensional fusion feature is first reduced to 50 dimensions using the principal component analysis method. Then, the 50-dimensional feature is standardized so that the mean of each feature is 0 and the variance is 1.

[0108] Optionally, based on the comprehensive spatio-temporal feature vector, the spatio-temporal evolution law of the fault outage event and its key influencing factors are identified, including:

[0109] The cause, impact degree, and recovery feature of the fault outage event are obtained.

[0110] Specifically, from a pre-constructed database, the cause, impact degree, and recovery feature of each fault outage event are obtained. These features are non-quantitative attributes of the event, for example, the cause can be "external damage" or "equipment aging", the impact degree can be divided into "mild" or "severe", and the recovery feature can be described as "fast recovery" or "long-time power outage". Through these attributes, it is intended to provide higher-dimensional analysis insights to fully represent the nature of the event.

[0111] Exemplarily, the cause of a certain fault outage event is "external damage", the impact degree is "moderate", and the recovery feature is "fast".

[0112] regularly mine the cause, influence degree and recovery feature with the comprehensive spatio-temporal feature vector;

[0113] Specifically, the cause, influence degree and recovery feature are regularly mined with the comprehensive spatio-temporal feature vector. The internal connection between the spatio-temporal feature of the event and the essential attribute of the event is explored. By calculating the correlation between the two, the mutual relationship is quantified, thereby revealing the deep connection between the seemingly independent event attribute and the spatio-temporal feature.

[0114] For example, the correlation degree between the "high incidence season" time feature of a certain area and the "lightning strike" fault cause is calculated. When the correlation coefficient reaches 0.8 or more, it indicates that there is a strong correlation between the two, and it can be inferred that lightning faults in the area are high in a certain season.

[0115] Based on the regular mining results, the evolution rules of different types of fault outage events in time and space are revealed, and the key factors affecting the events are identified.

[0116] Specifically, based on the regular mining results, the spatio-temporal features and event attributes with strong correlation are identified, and the evolution rules of different types of fault outage events in time and space are revealed. By matching, for example, the cause of "external damage" with the spatio-temporal feature of "high incidence in a certain area in summer afternoon", the key factors affecting the event can be accurately identified, and scientific basis for fault handling and prevention strategies can be provided.

[0117] For example, the regular mining results show that the outage events caused by "external damage" have a strong correlation with a specific "urban construction area" and "weekday daytime period", thereby revealing the spatio-temporal evolution rule of such faults and identifying "construction activities" as the key influencing factor.

[0118] Based on the same inventive concept, as shown in Figure 4 The present application also provides a spatio-temporal feature extraction system for multi-type fault outage events, which comprises:

[0119] An event monitoring and collection module is used to monitor the fault outage event signal in the power distribution network. When an event occurs, a collection process is triggered to collect multi-source fault outage event data of historical fault outage event data, device operation data and external environment data of the power distribution network, thereby obtaining an original data set.

[0120] A data cleaning module is used to clean the original data set based on the XGBoost model based on probability distribution, identify and remove noise and outliers, and complete missing data, thereby obtaining a standard data set.

[0121] A sample library construction module is used to construct a full-scenario fault outage sample library of the power distribution network based on the standard data set.

[0122] a time feature extraction module configured to extract time features of the fault outage event data based on the fault outage sample library by using a time convolution network algorithm, model and analyze periodicity and high-occurrence time periods of the fault outage event, and obtain a time feature vector;

[0123] a space feature extraction module configured to extract space features of the fault outage event data based on the fault outage sample library by using a graph convolution neural network algorithm, reveal geographical distribution rules and spatial correlations of different regions, devices and lines, and obtain a space feature vector;

[0124] a feature fusion module configured to dynamically weight and fuse the time feature vector and the space feature vector, and obtain a comprehensive space-time feature vector;

[0125] a rule mining module configured to identify space-time evolution rules of the fault outage event and key influence factors based on the comprehensive space-time feature vector.

[0126] It should be noted that the electrical connection between the above-mentioned units does not necessarily represent the direct connection of the line, the indirect connection mode, as long as the purpose of the application is achieved, which can be applied to the embodiments of the application. The above-mentioned is only an exemplary embodiment of the application, and cannot limit the scope of the application.

[0127] That is, any equivalent changes and modifications made according to the teachings of the present application are still within the scope of the present application. Other embodiments of the present application will be readily apparent to those skilled in the art upon considering the specification and practice of the true principles disclosed herein. The present application is intended to cover any variations, uses or adaptive changes of the present application that follow the general principles of the present application and include common knowledge or conventional technical means in the art not disclosed by the present application.

Claims

1. A method for extracting spatiotemporal features of multi-type fault outage events oriented to an event-driven architecture, characterized in that, The method includes: Monitor power outage event signals in the power distribution network, and trigger the data acquisition process when the event occurs; The data acquisition process involves collecting multi-source fault outage event data, including historical fault outage event data of the power distribution network, equipment operation data, and external environment data, to obtain the raw dataset. The original dataset is cleaned using an XGBoost model based on probability distribution to identify and remove noise and outliers, and to fill in missing data to obtain a standard dataset. Based on the aforementioned standard dataset, a full-scenario power outage sample library for power distribution networks will be constructed. Based on the aforementioned power outage sample database, a temporal convolutional network algorithm is used to extract time features from the power outage event data, model and analyze the periodicity and high-incidence periods of the power outage events, and obtain a time feature vector. Based on the aforementioned power outage sample database, spatial features of the power outage event data are extracted using a graph convolutional neural network algorithm to reveal the geographical distribution patterns and spatial correlations of different regions, equipment, and lines, thereby obtaining spatial feature vectors. The temporal and spatial feature vectors are dynamically weighted and fused to obtain a comprehensive spatiotemporal feature vector. This dynamic weighting and fusion includes: obtaining dynamic weight values, which reflect the influence of the temporal and spatial feature vectors on the power outage event; using the dynamic weight values ​​to perform a weighted summation of the temporal and spatial feature vectors to obtain a fused feature; and performing dimensionality reduction and standardization on the fused feature to output the comprehensive spatiotemporal feature vector. Based on the comprehensive spatiotemporal feature vector, the spatiotemporal evolution pattern and key influencing factors of the power outage event are identified. This identification includes: obtaining the causes, impact degree, and recovery characteristics of the power outage event; performing pattern mining on the causes, impact degree, and recovery characteristics with the comprehensive spatiotemporal feature vector; and based on the pattern mining results, revealing the temporal and spatial evolution patterns of different types of power outage events and identifying the key factors influencing the event.

2. The spatiotemporal feature extraction method for multiple types of power outage events oriented towards an event-driven architecture as described in claim 1, characterized in that, The monitored outage event signals in the power distribution network include: Set up an event listener to monitor the event logs and alarm signals in the power distribution network monitoring system and obtain event trigger signals; The event trigger signal is verified according to the verification rules used to distinguish between real faults and false alarms to obtain valid events; Based on the valid event, initialize the data acquisition parameters and trigger the data acquisition process.

3. The spatiotemporal feature extraction method for multiple types of power outage events oriented towards an event-driven architecture according to claim 1, characterized in that, The method further includes: optimizing the fault outage sample library, wherein: The aforementioned evolutionary patterns and key influencing factors are used to update the power outage sample database for all scenarios of the distribution network. Based on the updated fault outage sample database, the temporal convolutional network and graph convolutional neural network were retrained to optimize feature extraction performance. The feedback update and retraining steps are executed iteratively to achieve collaborative optimization of the fault outage sample library and the model.

4. The spatiotemporal feature extraction method for multiple types of power outage events oriented towards an event-driven architecture according to claim 1, characterized in that, The original dataset is cleaned using an XGBoost model based on probability distribution to identify and remove noise and outliers, and to complete missing data, including: The importance score of each feature in the original dataset is calculated using the XGBoost model to obtain a list of feature importance. Based on the feature importance list, features whose importance scores meet the predetermined filtering conditions are selected to form a subset of key features; The probability distribution of the key feature subset is estimated to generate a joint probability distribution model; Using the joint probability distribution model, noise and outliers in the original dataset are detected and removed, while missing data is filled in to output a standard dataset.

5. The spatiotemporal feature extraction method for multiple types of power outage events oriented towards an event-driven architecture according to claim 1, characterized in that, Based on the aforementioned power outage sample database, the temporal feature extraction of the power outage event data using a temporal convolutional network algorithm includes: Time series data, including power outage event sequences in daily, weekly, and seasonal dimensions, are obtained from the power outage sample library to obtain time series input. The time series input is processed by a temporal convolutional network to perform causal convolution and dilated convolution to capture long-term dependencies and obtain preliminary temporal features. The initial time features are adjusted using a threshold mechanism and optimized using residual connections to enhance feature stability and output a time feature vector.

6. The spatiotemporal feature extraction method for multiple types of power outage events oriented towards an event-driven architecture according to claim 1, characterized in that, Based on the aforementioned power outage sample database, spatial feature extraction of the power outage event data using a graph convolutional neural network algorithm includes: The power outage sample database is used to obtain distribution network topology data, including the connection relationships of stations, lines, transformers, and customers, to obtain the graph structure input; The graph convolutional neural network is used to aggregate the neighbor node features of the graph structure input, learn spatial correlation, and obtain preliminary spatial features. The preliminary spatial features are normalized and propagated in layers to expand the receptive field and output a spatial feature vector.

7. A spatiotemporal feature extraction system for multi-type power outage events oriented towards an event-driven architecture, characterized in that, The system is used for the spatiotemporal feature extraction method of multi-type power outage events oriented towards an event-driven architecture as described in any one of claims 1-6, and the system comprises: The event listening and acquisition module is used to listen for fault outage event signals in the distribution network. When an event occurs, it triggers the acquisition process to collect multi-source fault outage event data, including historical fault outage event data, equipment operation data, and external environment data of the distribution network, to obtain the raw dataset. The data cleaning module is used to perform XGBoost model cleaning based on probability distribution on the original dataset, identify and remove noise and outliers, fill in missing data, and obtain a standard dataset. The sample library construction module is used to construct a full-scenario power outage sample library for the power distribution network based on the standard dataset. The time feature extraction module is used to extract time features from the power outage event data based on the power outage sample library using a temporal convolutional network algorithm, model and analyze the periodicity and high-incidence periods of the power outage events, and obtain time feature vectors. The spatial feature extraction module is used to extract spatial features from the power outage event data based on the power outage sample library using a graph convolutional neural network algorithm, revealing the geographical distribution patterns and spatial correlations of different regions, equipment and lines, and obtaining spatial feature vectors. The feature fusion module is used to dynamically weight and fuse the temporal feature vector and the spatial feature vector to obtain a comprehensive spatiotemporal feature vector. The dynamic weighting and fusion of the temporal and spatial feature vectors includes: obtaining dynamic weight values, which reflect the degree of influence of the temporal and spatial feature vectors on the power outage event; using the dynamic weight values ​​to perform a weighted summation of the temporal and spatial feature vectors to obtain a fused feature; and performing dimensionality reduction and standardization on the fused feature to output the comprehensive spatiotemporal feature vector. The pattern mining module is used to identify the spatiotemporal evolution patterns and key influencing factors of the power outage event based on the comprehensive spatiotemporal feature vector. Identifying the spatiotemporal evolution patterns and key influencing factors of the power outage event includes: obtaining the causes, impact degree, and recovery characteristics of the power outage event; performing pattern mining on the causes, impact degree, and recovery characteristics with the comprehensive spatiotemporal feature vector; and based on the pattern mining results, revealing the temporal and spatial evolution patterns of different types of power outage events and identifying the key factors influencing the events.

Images