Power grid abnormal event early warning method, device and equipment and storage medium

Abstract

The application relates to a power grid abnormal event early warning method, device, equipment and storage medium. The method comprises the following steps: performing abnormal event detection on corresponding multi-source heterogeneous data in a key area of a power grid, determining an event data packet corresponding to at least one key trigger event occurring in the key area, inputting the event data packet into a small sample causal learning model based on a meta-learning framework to perform causal reasoning to output a causal data set of each key trigger event, performing association processing on the causal data set and a current dynamic causal knowledge graph of the power grid, obtaining a causal chain of each key trigger event, and determining an early warning report of each key trigger event according to each causal chain. The small sample causal learning model can improve the accuracy of the early warning report of the power grid abnormal event finally obtained.

Description

Technical Field

[0001] This application relates to the field of power grid technology, and in particular to a method, apparatus, device, and storage medium for early warning of abnormal power grid events. Background Technology

[0002] With the expansion of the power grid and the improvement of its intelligence level, the pressure on safety monitoring of key areas such as substations and transmission corridors in the power grid has increased dramatically.

[0003] In related technologies, the main approach is to use data streams from key areas, then analyze and process the data streams to generate early warning reports for abnormal events occurring in the key areas.

[0004] However, the relevant technologies suffer from the problem of low accuracy in the early warning reports of abnormal power grid events. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, apparatus, device, and storage medium for early warning of power grid anomalies that can improve the accuracy of early warning reports of acquired power grid anomalies, addressing the aforementioned technical problems.

[0006] Firstly, this application provides a method for early warning of abnormal events in the power grid, including:

[0007] Anomaly detection is performed based on multi-source heterogeneous data corresponding to key areas of the power grid to determine the event data packet corresponding to at least one key triggering event occurring in the key area.

[0008] The event data packets are input into a few-shot causal learning model based on a meta-learning framework for causal inference, and the causal datasets of each key triggering event are output. The few-shot causal learning model is used to learn the causal association information between different power grid events, including events that do not occur frequently in the power grid.

[0009] The causal datasets of each key triggering event are correlated with the current dynamic causal knowledge graph of the power grid to obtain the causal chain of each key triggering event;

[0010] Based on knowledge reasoning of each causal chain, early warning reports are determined for each key triggering event.

[0011] In one embodiment, abnormal event detection is performed based on multi-source heterogeneous data within a key area of ​​the power grid to determine the event data packet corresponding to at least one key triggering event occurring within the key area, including:

[0012] Event detection is performed based on multi-source heterogeneous data to obtain structured event information corresponding to at least one abnormal related event occurring in a key area;

[0013] When there are key triggering events among the various abnormal related events, an event data package corresponding to each key triggering event is constructed based on the structured event information and multi-source heterogeneous data of each abnormal related event.

[0014] In one embodiment, event detection is performed based on multi-source heterogeneous data to obtain at least one abnormal associated event occurring within a key area, including:

[0015] Time alignment is performed on multi-source heterogeneous data to obtain aligned heterogeneous data;

[0016] The aligned heterogeneous data is input into a cross-modal attention network model for power grid event detection, and the power grid event identifier of at least one power grid event occurring in the key area is output.

[0017] Based on a lightweight deep learning model, target detection is performed according to the power grid event identifier of each power grid event and the heterogeneous event data corresponding to each power grid event, so as to obtain the structured event information corresponding to the abnormal related events in each power grid event.

[0018] In one embodiment, the process of acquiring a dynamic causal knowledge graph includes:

[0019] Based on the causal datasets of each key triggering event, the initial causal knowledge graph of the power grid is updated to obtain a dynamic causal knowledge graph.

[0020] In one embodiment, the causal datasets of each key triggering event and the current dynamic causal knowledge graph of the power grid are correlated to obtain the causal chain of each key triggering event, including:

[0021] For any key triggering event, semantic matching is performed based on the causal dataset of the key triggering event and the current dynamic causal knowledge graph of the power grid to obtain the matching node in the dynamic causal knowledge graph;

[0022] Extract risk behavior chains associated with matching nodes from dynamic causal knowledge graphs as causal chains for key triggering events.

[0023] In one embodiment, knowledge reasoning is performed based on each causal chain to determine early warning reports for each key triggering event, including:

[0024] For any critical triggering event, a risk assessment is performed based on the causal chain of the critical triggering event, the preset risk severity of the terminal nodes in the causal chain, the corresponding link confidence and link complexity, to obtain the risk level of the causal chain;

[0025] The causal chain and the corresponding risk level are input into a pre-built large language model for enhanced reasoning and language interpretation, resulting in early warning reports for key triggering events.

[0026] In one embodiment, the method further includes:

[0027] Warning reports for each critical triggering event are sent to the user terminal, instructing the user to take security measures for each critical triggering event.

[0028] Secondly, this application also provides a power grid abnormal event early warning device, comprising:

[0029] The detection module is used to detect abnormal events based on multi-source heterogeneous data corresponding to key areas of the power grid, and to determine the event data packet corresponding to at least one key triggering event occurring in the key area.

[0030] The causal reasoning module is used to input event data packets into a few-shot causal learning model based on a meta-learning framework for causal reasoning, and output causal datasets of key triggering events. The few-shot causal learning model is used to learn causal association information between different power grid events, including events that do not occur frequently in the power grid.

[0031] The correlation processing module is used to correlate the causal datasets of each key triggering event with the current dynamic causal knowledge graph of the power grid to obtain the causal chain of each key triggering event.

[0032] The knowledge reasoning module is used to perform knowledge reasoning based on each causal chain and determine the early warning reports for each key triggering event.

[0033] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method in any of the embodiments of the first aspect described above.

[0034] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method in any of the embodiments of the first aspect described above.

[0035] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method in any of the embodiments of the first aspect described above.

[0036] The aforementioned power grid anomaly event early warning method, device, equipment, and storage medium include: detecting anomalies based on multi-source heterogeneous data corresponding to key areas of the power grid; determining event data packets corresponding to at least one key triggering event occurring in the key area; inputting the event data packets into a few-shot causal learning model based on a meta-learning framework for causal inference; outputting causal datasets for each key triggering event; performing correlation processing between the causal datasets for each key triggering event and the current dynamic causal knowledge graph of the power grid to obtain causal chains for each key triggering event; and determining early warning reports for each key triggering event based on knowledge inference using each causal chain. The few-shot causal learning model is used to learn causal association information between different power grid events, including events that do not frequently occur in the power grid. The aforementioned method employs a small-sample causal learning model to perform causal inference on rare or small-sample events that do not occur frequently in the power grid. This broadens the applicability of the method and improves the feasibility of causal inference, avoiding the problem of missed inferences in rare and high-risk scenarios, thereby greatly improving the accuracy of causal inference. Simultaneously, this method only processes data packets of detected key triggering events or abnormal events (not the entire dataset), allowing for on-demand processing and avoiding processing of data from risk-free periods, reducing computational effort. This not only accelerates the speed of abnormal event warnings and improves their timeliness and efficiency but also reduces computational errors in complex scenarios, improving the accuracy of causal inference and further enhancing the accuracy of the final power grid abnormal event warning reports. Furthermore, based on key triggering events (i.e., power grid abnormal events), the method integrates multi-source heterogeneous data and achieves abnormal event tracing and warning through causal inference, providing interpretable, traceable, and quantifiable decision-making basis for power safety management. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart illustrating a power grid anomaly event early warning method in one embodiment;

[0039] Figure 2 This is a flowchart illustrating a power grid anomaly event early warning method in another embodiment;

[0040] Figure 3 This is a flowchart illustrating a power grid anomaly event early warning method in another embodiment;

[0041] Figure 4 This is a flowchart illustrating a power grid anomaly event early warning method in another embodiment;

[0042] Figure 5 This is a flowchart illustrating a power grid anomaly event early warning method in another embodiment;

[0043] Figure 6 This is a structural block diagram of a power grid abnormal event early warning device in one embodiment;

[0044] Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0046] The power grid anomaly event early warning method provided in this application embodiment can be applied to power safety production monitoring scenarios, specifically to a power grid anomaly event early warning system. This system includes computer equipment and a power grid data acquisition system. The computer equipment can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and IoT devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart vehicle devices, projection devices, servers, etc. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. Simultaneously, the data acquisition system can include video acquisition equipment, audio acquisition equipment, and device remote signaling data acquisition equipment.

[0047] The computer equipment communicates with each acquisition device in the data acquisition system via various methods, such as Bluetooth, mobile data, wired connection, or Wi-Fi. In practical applications, the computer equipment receives the corresponding data collected by each acquisition device in the data acquisition system.

[0048] In one exemplary embodiment, such as Figure 1 As shown, a method for early warning of abnormal events in a power grid is provided. Taking the application of this method to a computer device in a power grid abnormal event early warning system as an example, the method includes the following steps S201 to S204. Wherein:

[0049] Step S201: Perform abnormal event detection based on the multi-source heterogeneous data corresponding to the key area of ​​the power grid, and determine the event data packet corresponding to at least one key triggering event occurring in the key area.

[0050] The aforementioned power grid can be of various types, including radial power grid, ring power grid, mesh power grid, AC power grid, DC power grid, distribution network, and transmission network, etc., and this application embodiment does not limit this type. Optionally, the key area of ​​the power grid can be the core grid area composed of substations, transmission corridors, switching stations, transmission lines, etc.

[0051] In practical applications, multi-source heterogeneous data can include at least two of the following: telemetry data, geographic data, equipment status data, remote sensing data, log data, etc. However, in this embodiment, multi-source heterogeneous data including video data, audio data, and equipment remote signaling data (such as switch position, protection device action signal, etc.) are used as an example for illustration.

[0052] It should be noted that the video data in multi-source heterogeneous data can include multiple key video frames. The computer device can decode the video stream acquired by the video acquisition device in real time to obtain multiple raw video frames, and then extract multiple key video frames as needed or at a fixed frame rate to obtain the video data in the multi-source heterogeneous data.

[0053] Meanwhile, audio data in multi-source heterogeneous data can represent abnormal sounds generated in key areas of the power grid (such as discharge sounds, abnormal equipment noises, etc.); remote signaling data of equipment (such as circuit breakers, disconnect switches, protection devices, etc.) in multi-source heterogeneous data can represent switching signals of equipment operating status, such as "open or closed" position, "alarm or reset" status, etc.

[0054] Specifically, computer equipment can employ anomaly detection algorithms to detect anomalies in multi-source heterogeneous data within key areas of the power grid, and identify event data packets corresponding to at least one key triggering event occurring within the key area.

[0055] In addition, the computer equipment can pre-train an abnormal event detection model, and then input the corresponding multi-source heterogeneous data in the key area of ​​the power grid into the abnormal event detection model for abnormal event detection, and output the event data packet corresponding to at least one key triggering event occurring in the key area.

[0056] Power grid events occurring within critical areas of the power grid can include both normal and abnormal events. It should be noted that the aforementioned critical triggering events are typically power grid events that may indicate the onset or escalation of risk, such as "a safety helmet disappearing" or "entering an electronic fence," and may themselves be abnormal events.

[0057] In practical applications, a time window is constructed centered on the time period of the key triggering event, extending a period of time to both sides of the center. All relevant data (including corresponding video data, audio data, and device remote signaling data) within this time window from multi-source heterogeneous data is acquired. Simultaneously, power grid events within this time window can be detected, and their power grid event information can be obtained. Furthermore, all relevant data within this time window and the power grid event information can be combined to form structured event data for the key triggering event. Finally, the structured event data for each key triggering event can be packaged to generate event data packets corresponding to each key triggering event. Optionally, the aforementioned structured event data can be structured JSON data.

[0058] It should be noted that the event data packet corresponding to each key triggering event can include all the context data required for subsequent risk analysis of each key triggering event.

[0059] In this embodiment, the number of key triggering events corresponding to the event data packet can be less than or equal to a preset threshold N, or the number of key triggering events corresponding to the event data packet can be equal to the number of key triggering events corresponding to a specific combination of key triggering events. For example, the "safety helmet disappears" event is followed immediately by the "person climbing" event. These two events can be called a specific combination of key triggering events, and the number of key triggering events corresponding to this specific combination of key triggering events is equal to 2.

[0060] In practical applications, the power grid can be monitored for safety in real time, and event data packets can be continuously constructed. This application embodiment uses an event data packet as an example for illustration.

[0061] Step S202 involves inputting the event data packet into a few-shot causal learning model based on a meta-learning framework for causal inference, outputting a causal dataset of each key triggering event. The few-shot causal learning model is used to learn the causal relationships between different power grid events, including those that are infrequent in the power grid.

[0062] In practical applications, computer devices can input event data packets into a few-sample causal learning model based on a meta-learning framework for causal inference, and then output causal datasets corresponding to each key triggering event occurring in the key region.

[0063] It should be noted that, in response to the problem of extremely scarce samples of serious violations occurring in the power grid, this application introduces a metric meta-learning framework based on "prototype networks" to construct a small-sample causal learning model based on the meta-learning framework. The core of this small-sample causal learning model based on the meta-learning framework is to extract features from event data packets to obtain the prototype vectors of each key triggering event, and to calculate the distance between the prototype vectors of each key triggering event and the normal prototype vectors of the standard events corresponding to each key triggering event to obtain the causal deviation between each key triggering event and the corresponding standard event.

[0064] The prototype vector of a critical triggering event represents the operational behavior and state changes associated with that event. The standard event corresponding to a critical triggering event can be understood as a normal event where workers perform tasks according to procedures, processes, and specifications, with the operation sequence, actions, timing, location, and equipment status all meeting safety requirements, and without violations, dangerous actions, or unauthorized operations. This is the opposite of the critical triggering event. Simultaneously, the standard event corresponding to a critical triggering event can be understood as the degree of deviation of the critical triggering event from its corresponding standard event in terms of operational sequence, execution logic, and causal dependencies.

[0065] The few-shot causal learning model based on the meta-learning framework can be understood as a few-shot learning model built on the meta-learning architecture and combined with a causal inference mechanism. In this embodiment, the few-shot causal learning model is used to mine causal association information between different power grid events. Power grid events include infrequent few-shot abnormal events or serious violations in the power grid (such as rare equipment failures, sudden external damage, and unconventional operation events caused by extreme weather). When applied, this few-shot causal learning model can learn the hidden causal structure and causal features in multi-source heterogeneous data (video data, audio data, and equipment remote signaling data), establish causal dependencies, propagation paths, and impact mechanisms between different power grid events, and realize causal inference, cause tracing, and type identification of power grid events (including few-shot power grid events) occurring in key areas of the power grid.

[0066] The training process of the meta-learning framework is explained here: a multi-source heterogeneous dataset corresponding to all standard events in the power grid area can be obtained, and a training set and a validation set can be constructed by sampling from the multi-source heterogeneous dataset. Then, the meta-learning framework is trained based on the training set and the validation set, so that the meta-learning framework can extract the deep features that best represent its causal pattern from the input multi-source heterogeneous data, and distinguish different key triggering events based on the deep features.

[0067] In practical applications, the meta-learning framework can also analyze the temporal characteristics of the corresponding events in the input, such as the event sequence and time interval, and compare them with the temporal characteristics of the standard events corresponding to the event to identify logical error information of logical error events such as reversed order and missing key steps.

[0068] In the embodiments of this application, the causal dataset of each key triggering event output by the few-shot causal learning model may include the causal deviation between each key triggering event and the corresponding standard event, the logical error information of each key triggering event, the abnormal confidence of each key triggering event, and the event data packet of each key triggering event.

[0069] It should be noted that, over time, the training and validation sets can be updated to continuously improve the few-shot causal learning model and enhance its processing accuracy.

[0070] Step S203: Perform correlation processing on the causal datasets of each key triggering event and the current dynamic causal knowledge graph of the power grid to obtain the causal chain of each key triggering event.

[0071] Specifically, for each key triggering event, the computer device can employ a data association algorithm to associate the causal dataset of the key triggering event with the current dynamic causal knowledge graph of the power grid, thereby obtaining the causal chain of the key triggering event. Optionally, the aforementioned data association algorithm can be at least one of statistical or rule-based association algorithms, feature fusion association algorithms, probabilistic reasoning association algorithms, etc.

[0072] Simultaneously, the computer device can also acquire a pre-built correlation processing model, and then input the causal dataset of the key triggering event and the current dynamic causal knowledge graph of the power grid into the correlation processing model for correlation processing, and output the causal chain of the key triggering event.

[0073] It should be noted that the current dynamic causal knowledge graph of the power grid can be understood as the current real causal knowledge graph of the power grid; in this embodiment of the application, in order to reduce computing power, the current dynamic causal knowledge graph of the power grid can be the current dynamic causal knowledge graph of the key areas in the power grid.

[0074] Step S204: Perform knowledge reasoning based on each causal chain to determine the early warning report for each key triggering event.

[0075] Specifically, for any key triggering event, the computer device can pre-train a knowledge reasoning model, then input the causal chain of the key triggering event into the knowledge reasoning model for knowledge reasoning, and output an early warning report for the key triggering event. Optionally, the above-mentioned knowledge reasoning model can be implemented by at least one of the following: convolutional neural network model, fully connected neural network model, residual neural network model, long short-term memory neural network model, etc.

[0076] In practical applications, computer devices can also employ knowledge reasoning algorithms to perform knowledge reasoning based on the causal chain of the key triggering event, thereby obtaining an early warning report for the key triggering event. Optionally, the aforementioned knowledge reasoning algorithm can be a representation learning-based reasoning algorithm, a hybrid reasoning algorithm, or an ontology-based reasoning algorithm, etc.

[0077] In practical applications, it is necessary to safely handle critical triggering events occurring in key areas of the power grid. The implementation method of this process is described below. In one embodiment, after executing the process in step S204 above, the method may further include: sending early warning reports of each critical triggering event to the user terminal, instructing that each critical triggering event be handled safely.

[0078] Specifically, computer equipment can send early warning reports of each critical triggering event to user terminals to notify on-site users and the power grid's remote monitoring center, instructing staff to handle each critical triggering event safely.

[0079] In this embodiment, the information in the aforementioned warning report is in natural language and is interpretable, thus making the key triggering events actionable. In practical applications, users can take safe measures for each key triggering event based on the warning report displayed on the user terminal. Meanwhile, the aforementioned user terminal can be a display device such as an audible and visual alarm, a monitoring terminal, or a mobile terminal.

[0080] The technical solution in this application embodiment detects abnormal events based on multi-source heterogeneous data corresponding to key areas of the power grid, determines the event data packet corresponding to at least one key triggering event occurring in the key area, inputs the event data packet into a few-shot causal learning model based on a meta-learning framework for causal inference, outputs causal datasets for each key triggering event, correlates the causal datasets of each key triggering event with the current dynamic causal knowledge graph of the power grid, obtains the causal chain of each key triggering event, and determines the early warning report for each key triggering event based on knowledge inference of each causal chain. The few-shot causal learning model is used to learn the causal association information between different power grid events, including events that do not frequently occur in the power grid. The above method uses a few-shot causal learning model to perform causal inference on scarce events or few-shot events that do not frequently occur in the power grid, making the method... This method has a wider range of applications and can improve the feasibility of causal inference, avoiding the problem of missed inference in rare and high-risk scenarios, thus greatly improving the accuracy of causal inference. At the same time, this method only processes data packets of detected key triggering events or abnormal events (not all data packets), which allows for on-demand processing and avoids processing data during risk-free periods, reducing computing power. This not only speeds up the warning of abnormal events and improves the timeliness and efficiency of warnings, but also reduces computational errors in complex scenarios, improving the accuracy of causal inference and further enhancing the accuracy of the final warning reports of abnormal power grid events. In addition, the above method takes key triggering events (i.e., abnormal power grid events) as a premise, integrates multi-source heterogeneous data, and achieves abnormal event tracing and warning through causal inference, providing interpretable, traceable, and quantifiable decision-making basis for power safety management.

[0081] The following describes the process of anomaly detection based on multi-source heterogeneous data within a key area of ​​the power grid. In one embodiment, as... Figure 2 As shown, the process in step S201 above can be implemented in the following way:

[0082] Step S211: Perform event detection based on multi-source heterogeneous data to obtain structured event information corresponding to at least one abnormal related event occurring in the key area.

[0083] The computer equipment can employ a multi-source event fusion detection algorithm to detect events based on multi-source heterogeneous data, obtaining structured event information corresponding to at least one abnormal related event occurring within a key area. Optionally, the multi-source event fusion detection algorithm can be at least one of feature-level fusion algorithms, decision-level fusion algorithms, event detection algorithms, and event differentiation algorithms.

[0084] Step S212: If there are key triggering events in each abnormal associated event, construct the event data packet corresponding to each key triggering event based on the structured event information and multi-source heterogeneous data of each abnormal associated event.

[0085] Furthermore, the computer device can detect whether there is at least one key triggering event among the various abnormal related events. If so, it constructs an event data packet corresponding to each key triggering event based on the structured event information and multi-source heterogeneous data of each abnormal related event.

[0086] One method for detecting whether there is at least one key triggering event among each abnormal associated event is to obtain the event identifier of each abnormal associated event, then match the event identifier of each abnormal associated event with the event identifier in the key triggering event identifier list, obtain the successfully matched event identifier, and determine the event corresponding to the successfully matched event identifier as the key triggering event.

[0087] Meanwhile, another way to detect whether there is at least one key triggering event among the various abnormal related events is to obtain a pre-trained algorithm model, then input each abnormal related event into the algorithm model, and the algorithm model outputs the detection result of whether there is at least one key triggering event among each abnormal related event.

[0088] In addition, the method of constructing event data packets corresponding to each key triggering event based on the structured event information of each abnormal related event and the multi-source heterogeneous data can be to extract the heterogeneous data corresponding to each key triggering event from the multi-source heterogeneous data according to a preset matching method, and then fuse these heterogeneous data with the structured event information of each abnormal related event to obtain the related event information of each abnormal related event. Then, the related event information of each abnormal related event is combined to generate an event data packet.

[0089] In one embodiment, such as Figure 3 As shown, the process in step S211 above, which involves obtaining structured event information corresponding to at least one abnormal related event occurring within a key area through event detection based on multi-source heterogeneous data, can be implemented in the following way:

[0090] Step S2111: Time alignment is performed based on the multi-source heterogeneous data to obtain aligned heterogeneous data.

[0091] In this embodiment of the application, the multi-source heterogeneous data includes video data, audio data and device remote signaling data within the key area, and all three types of data carry timestamps.

[0092] Specifically, computer equipment can use the Network Time Protocol (NTP) or a hardware clock synchronization mechanism to perform high-precision time alignment based on the timestamps in the multi-source heterogeneous data, thereby obtaining aligned heterogeneous data to ensure that the video data, audio data, and device remote signaling data in the aligned heterogeneous data are under the same time reference.

[0093] In this embodiment, to improve the accuracy of abnormal event warnings and provide stable input for subsequent processing, the computer device can first preprocess the multi-source heterogeneous data to obtain preprocessed heterogeneous data, and then perform time alignment based on the timestamps in the preprocessed heterogeneous data to obtain aligned heterogeneous data. The preprocessing methods may include sampling, noise reduction, white balance correction, and resolution normalization. Here, sampling can be understood as sampling keyframes or low-frequency frames from the preprocessed heterogeneous data to reduce computational power and lower the complexity of causal inference.

[0094] Step S2112: Input the aligned heterogeneous data into the cross-modal attention network model for power grid event detection, and output the power grid event identifier of at least one power grid event occurring in the key area.

[0095] Optionally, the aforementioned cross-modal attention network model can consist of a multimodal feature encoding sub-network, a feature projection layer, a cross-modal attention layer, a feedforward network layer, a multi-layer cross-modal attention stacked layer, a global feature aggregation layer, and a detection output layer. In practical applications, the aforementioned cross-modal attention network model can first fuse the aligned heterogeneous data, and then perform power grid event detection processing based on the fused data.

[0096] Specifically, the computer device can input aligned heterogeneous data into a cross-modal attention network model for power grid event detection, and then output a power grid event identifier for at least one power grid event occurring in a key area.

[0097] Step S2113: Based on a lightweight deep learning model, target detection is performed according to the power grid event identifier of each power grid event and the heterogeneous event data corresponding to each power grid event to obtain the structured event information corresponding to the abnormal related events in each power grid event.

[0098] The aforementioned lightweight deep learning model can consist of a lightweight feature extraction module (such as a depthwise separable convolutional network, a pointwise convolutional network, a Ghost convolutional network, or a small kernel convolutional network), a lightweight attention module (such as a lightweight channel attention network, a spatial attention network, or a lightweight cross-modal attention network), a multi-scale feature fusion module, and a detection output module (such as a lightweight fully connected layer or a global pooling layer). In this embodiment, the lightweight deep learning model consists of a lightweight YOLOv8n model and a Temporal Segment Network (TSN) model.

[0099] It should be noted that the aforementioned abnormal associated events can include visual events (such as "safety helmet", "open flame", "personnel intrusion", etc.), audio events (such as "abnormal discharge sound", etc.), and state change events (such as "switch tripping", etc.). In practical applications, abnormal associated events are either abnormal events or ambiguous events (which may be normal events), for example, "not wearing a safety helmet" (abnormal event) and "operator approaching the equipment" (possibly a normal event).

[0100] In practical applications, computer equipment can input the power grid event identifiers of each power grid event and the heterogeneous event data corresponding to each power grid event into a lightweight deep learning model for target detection, and then output the structured event information corresponding to the abnormal related events in each power grid event.

[0101] Optionally, the structured event information corresponding to the aforementioned abnormal associated events may include information such as the event type, timestamp and spatial location of the occurrence, and confidence level of the abnormal associated events.

[0102] The technical solution in this application embodiment performs event detection based on multi-source heterogeneous data to obtain structured event information corresponding to at least one abnormal associated event occurring in a key area. When key triggering events exist among the abnormal associated events, event data packets corresponding to each key triggering event are constructed based on the structured event information of each abnormal associated event and the multi-source heterogeneous data. This method can acquire half the amount of data as needed, i.e., construct event data packets corresponding to each key triggering event, thereby reducing the data volume for subsequent abnormal event warning processing, accelerating the speed of subsequent abnormal event warnings, and improving the timeliness of abnormal event intervention. Furthermore, with a smaller data volume, it can also reduce subsequent processing errors, thereby improving the accuracy of warnings. In addition, the above method can use multi-source heterogeneous data as input for abnormal event warning processing, which can effectively suppress interference caused by changes in illumination, background noise, etc., in single-dimensional data, solving the problem of false warnings caused by insufficient reliability of single-dimensional data, and improving the accuracy of abnormal event warnings.

[0103] In one example, the process of obtaining the dynamic causal knowledge graph mentioned above may include: updating the initial causal knowledge graph of the power grid based on the causal dataset of each key triggering event to obtain the dynamic causal knowledge graph.

[0104] Specifically, the computer equipment can update the initial causal knowledge graph of the power grid according to the preset graph update rules and the causal dataset of each key triggering event, so as to obtain a dynamic causal knowledge graph.

[0105] In addition, the computer device can also acquire a pre-built graph update model, and then input the causal datasets of each key triggering event and the initial causal knowledge graph of the power grid into the graph update model to update the initial causal knowledge graph, and output the current dynamic causal knowledge graph of the power grid.

[0106] It's important to note that after each power grid anomaly warning, the previous dynamic causal knowledge graph can be updated to generate a new one, ready for use when issuing the next power grid anomaly warning. In other words, the dynamic causal knowledge graph can be iteratively updated. Updates may include correcting rule weights, adding new causal edges, etc.

[0107] The structure of a dynamic causal knowledge graph is described here: A dynamic causal knowledge graph adopts a three-layer node system of entity-event-risk and a two-layer edge relationship system of causality-spatiotemporal relationships. Optionally, a dynamic causal knowledge graph consists of five parts: nodes, edges, attributes, causal paths, and knowledge rules.

[0108] Among them, nodes are the basic units of dynamic causal knowledge graphs, covering all objects involved in causal relationships in key areas of the power grid. They are classified by type as follows: event nodes (various power grid events occurring in key areas, including causal events, process events, and result events, such as "wildfire outbreak", "tower tilting", "line tripping", "voltage exceeding limit", etc.), equipment nodes (core power grid equipment in key areas, such as "±800kV converter station", "500kV hub substation", "ultra-high voltage transmission line", "protection device", etc.), environmental nodes (external environmental factors affecting power grid operation, such as "rainstorm", "typhoon", "icing", "high temperature", "wildfire", etc.), and parameter nodes (key indicators characterizing the power grid operating status, such as "current", "voltage", "oil temperature", "partial discharge value", "power flow density", etc.).

[0109] In a dynamic causal knowledge graph, edges represent directed causal relationships and association types. Edges are directed connections between nodes and are used to clarify the logical and topological relationships between nodes. They mainly include: causal edges (representing causal relationships between events such as induction, cause, aggravation, and mitigation, with a clear direction (from cause to effect), such as "wildfire outbreak → air insulation breakdown" and "oil temperature exceeding the standard → transformer failure"), and association edges (representing physical connections, subordinate relationships, or attribute associations between entities, providing topological support for causal reasoning, such as "converter station - connection - ultra-high voltage line", "transmission corridor - coverage - geographical area", and "equipment - monitoring - partial discharge value").

[0110] In a dynamic causal knowledge graph, attributes represent key information for quantifying and constraining causal relationships. Attributes are additional descriptive information for nodes and edges, providing a quantitative basis for risk assessment and knowledge reasoning. Attributes include: node attributes (including event occurrence timestamps, event types, equipment voltage levels, years of operation, normal threshold ranges of parameters, measured values, etc.) and edge attributes (including causal strength (confidence level), triggering conditions, connection methods of associated edges, influence weights, etc.).

[0111] In a dynamic causal knowledge graph, causal paths can represent the complete logical chain of event evolution. Each causal path consists of multiple causal edges connected sequentially, describing the complete evolution of an abnormal event from its cause to its consequence. Example: Typhoon (environmental node) → Tower foundation scour (event node) → Tower tilting (event node) → Conductor-to-ground discharge (event node) → Line tripping (event node) → Power outage at load center (event node). Each path includes attributes such as total causal strength, evolution duration, and the range of affected equipment.

[0112] The knowledge rules in the dynamic causal knowledge graph serve as prior constraints for causal reasoning. These knowledge rules are explicit constraints based on knowledge and operational procedures within the power grid domain, embedded in the dynamic causal knowledge graph to aid reasoning and judgment under small-sample anomaly events. These knowledge rules can include causal rules (e.g., "If the ice thickness on the transmission line is >20mm and the wind speed is >8m / s, then the confidence level of conductor galloping is ≥0.9") and constraint rules (e.g., "Oil temperature >95℃ in the hub substation is a necessary triggering condition for transformer overload").

[0113] The technical solution in this application embodiment updates the initial causal knowledge graph of the power grid based on the causal dataset of each key triggering event to obtain a dynamic causal knowledge graph. The above method can continuously update the causal knowledge graph of the power grid to improve the accuracy of the causal knowledge graph used in subsequent processing, thereby improving the accuracy of causal reasoning.

[0114] The following describes the steps for associating the causal datasets of each key triggering event with the current dynamic causal knowledge graph of the power grid. In one embodiment, as... Figure 4 As shown, the process in step S203 above may include:

[0115] Step S213: For any key triggering event, perform semantic matching based on the causal dataset of the key triggering event and the current dynamic causal knowledge graph of the power grid to obtain the matching node in the dynamic causal knowledge graph.

[0116] In this embodiment of the application, the dynamic causal knowledge graph can be used to search for risk transmission paths of key triggering events and assess risk levels.

[0117] Specifically, the computer device can employ a semantic matching algorithm to perform semantic matching or entity alignment based on the causal dataset of key triggering events and the current dynamic causal knowledge graph of the power grid, thereby obtaining matching nodes in the dynamic causal knowledge graph. Optionally, the aforementioned semantic matching algorithm can be at least one of the following: basic string matching algorithm, deep semantic representation matching algorithm, graph-specific matching algorithm, lightweight matching algorithm, etc.

[0118] In this embodiment of the application, the computer device can determine the event type, event attribute, or event name of the key triggering event based on the causal dataset of the key triggering event, and then perform semantic matching between the event type of the key triggering event and the event type, event attribute, or event name of each node in the current dynamic causal knowledge graph of the power grid to obtain the matching node in the dynamic causal knowledge graph.

[0119] Step S223: Extract the risk behavior chain associated with the matching node from the dynamic causal knowledge graph as the causal chain of the key triggering event.

[0120] In this embodiment of the application, the computer device can extract the risk behavior chain associated with the matching node as the causal chain of the key triggering event from the dynamic causal knowledge graph. The risk behavior chain associated with the matching node can be formed by connecting all the nodes associated with the matching node and the edges between these nodes. It is usually a directed path composed of nodes and edges, such as: "person not wearing a safety helmet" → (entering) → "high-altitude work platform" → (causing) → "high risk of falling from height".

[0121] It should be noted that the above causal chain can include events, entities, risk levels, causal relationships, and confidence levels.

[0122] The technical solution in this application embodiment, for any key triggering event, performs semantic matching based on the causal dataset of the key triggering event and the current dynamic causal knowledge graph of the power grid to obtain matching nodes in the dynamic causal knowledge graph, and extracts the risk behavior chain associated with the matching node from the dynamic causal knowledge graph as the causal chain of the key triggering event; the above method can utilize the latest dynamic causal knowledge graph of the power grid to process complex and sudden key triggering events to obtain the corresponding causal chain, and on this basis, can improve the early warning capability of special abnormal events.

[0123] The process of knowledge reasoning based on each causal chain described above is explained below. In one embodiment, such as... Figure 5 As shown, the process in step S204 above can be implemented in the following way:

[0124] Step S214: For any critical triggering event, risk assessment is performed based on the causal chain of the critical triggering event, the preset risk severity of the end node in the causal chain, the corresponding link confidence and link complexity, to obtain the risk level of the causal chain.

[0125] In practical applications, for any critical triggering event, a computer device can employ a risk assessment algorithm to evaluate the risk level of the causal chain based on the causal chain of the critical triggering event, the preset risk severity of the terminal node in the causal chain, the corresponding link confidence, and the link complexity. In this embodiment, the preset risk severity of the terminal node in the causal chain is determined based on all nodes preceding the terminal node; the link confidence of the causal chain can be understood as the probability that the causal chain will generate risk; and the risk level is interpretable.

[0126] Optionally, the aforementioned risk assessment algorithm may include one or more of the following: probability and statistics-based assessment algorithm, fuzzy logic assessment algorithm, multi-attribute decision assessment algorithm, and lightweight risk assessment algorithm.

[0127] In this embodiment of the application, the computer device may use a weighted or rule-based mapping method to perform risk assessment based on the causal chain of the key triggering event, the preset risk severity of the end node in the causal chain, the corresponding link confidence and link complexity.

[0128] Step S224: Input the causal chain and the corresponding risk level into the pre-built large language model for enhanced reasoning and language interpretation to obtain the early warning report corresponding to the key triggering event.

[0129] Furthermore, computer devices can input causal chains and corresponding risk levels into a pre-built large language model for enhanced reasoning and language interpretation, thereby obtaining early warning reports corresponding to key triggering events.

[0130] Among them, enhanced reasoning and language interpretation can be understood as transforming the deep causal reasoning results in the causal chain into natural language early warning information that conforms to the context of power safety regulations, includes risk description, cause tracing, and specific handling suggestions, so as to obtain actionable and interpretable early warning reports (natural language text, such as: "Warning! Personnel A was detected working at height in area XX without wearing a safety helmet, posing a risk of falling from height. Please stop them immediately and require them to wear the correct safety protective equipment"), and based on the early warning report, key triggering events can be directly addressed.

[0131] In this embodiment, the aforementioned large language model is constructed by adaptively fine-tuning a general-purpose large language model using massive amounts of professional corpora from the power grid field, such as power safety regulations documents, historical power grid event reports, and expert response plans. This results in a model with text understanding and generation capabilities. Furthermore, the final fine-tuned large language model deeply internalizes the professional logic, terminology, and risk context of power safety.

[0132] In practical applications, the above-mentioned large language model can perform the following steps when performing enhanced reasoning and language interpretation: ① restate and refine the causal chain to generate a risk description that conforms to human reading habits; ② perform secondary reasoning based on the risk description and the corresponding risk level, such as citing relevant specific safety specification clause numbers and contents; ③ further combine best practices in the power grid industry to generate specific, step-by-step on-site handling suggestions and subsequent preventive measures, and generate early warning reports corresponding to key triggering events based on this information.

[0133] Furthermore, the results of the handling can be obtained, and the key triggering events, causal chains, early warning reports, and handling results can be structured and stored in the database as reference data for subsequent applications.

[0134] The technical solution in this application embodiment, for any critical triggering event, performs a risk assessment based on the causal chain of the critical triggering event, the preset risk severity of the terminal nodes in the causal chain, the corresponding link confidence and link complexity, and obtains the risk level of the causal chain. The causal chain and the corresponding risk level are then input into a pre-built large language model for enhanced reasoning and language interpretation to obtain a warning report corresponding to the critical triggering event. The above method can use a large language model to deeply understand and reason about the risks of critical triggering events occurring in key areas of the power grid. This not only makes the final warning report interpretable and transparent, greatly improving the practical value of the warning information and the response efficiency of safety monitoring personnel, but also enables the power grid system to achieve a leap or closed loop from "perceiving an anomaly" to "understanding the risk" and then to "predicting the consequences".

[0135] To facilitate understanding by those skilled in the art, the following provides a detailed description of the power grid abnormal event early warning method provided in this application. This method may include:

[0136] Step S601: Perform time alignment based on the multi-source heterogeneous data to obtain aligned heterogeneous data;

[0137] Step S602: Input the aligned heterogeneous data into the cross-modal attention network model for power grid event detection, and output the power grid event identifier of at least one power grid event occurring in the key area;

[0138] Step S603: Based on a lightweight deep learning model, target detection is performed according to the power grid event identifier of each power grid event and the heterogeneous event data corresponding to each power grid event to obtain the structured event information corresponding to the abnormal related events in each power grid event.

[0139] Step S604: If there are key triggering events in each abnormal associated event, construct the event data package corresponding to each key triggering event based on the structured event information and multi-source heterogeneous data of each abnormal associated event;

[0140] Step S605: Input the event data packet into the few-shot causal learning model based on the meta-learning framework for causal inference, and output the causal dataset of each key triggering event; the few-shot causal learning model is used to learn the causal relationship information between different power grid events, including events that do not occur frequently in the power grid;

[0141] Step S606: For any key triggering event, perform semantic matching based on the causal dataset of the key triggering event and the current dynamic causal knowledge graph of the power grid to obtain the matching node in the dynamic causal knowledge graph.

[0142] Step S607: Extract the risk behavior chain associated with the matching node from the dynamic causal knowledge graph as the causal chain of the key triggering event; the dynamic causal knowledge graph is obtained by updating the initial causal knowledge graph of the power grid based on the causal dataset of each key triggering event.

[0143] Step S608: For any critical triggering event, risk assessment is performed based on the causal chain of the critical triggering event, the preset risk severity of the end node in the causal chain, the corresponding link confidence and link complexity, to obtain the risk level of the causal chain.

[0144] Step S609: Input the causal chain and the corresponding risk level into the pre-built large language model for enhanced reasoning and language interpretation to obtain the early warning report corresponding to the key triggering event;

[0145] Step S610: Send the early warning reports of each critical triggering event to the user terminal, instructing that each critical triggering event be handled securely.

[0146] It should be noted that the descriptions of steps S601-S610 above can be found in the relevant descriptions in the above embodiments, and their effects are similar, so they will not be repeated here.

[0147] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0148] Based on the same inventive concept, this application also provides a power grid anomaly event early warning device for implementing the aforementioned power grid anomaly event early warning method. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more power grid anomaly event early warning device embodiments provided below can be found in the limitations of the power grid anomaly event early warning method described above, and will not be repeated here.

[0149] In one embodiment, such as Figure 6 As shown, a power grid abnormal event early warning device is provided, including: a detection module 11, a causal reasoning module 12, an association processing module 13, and a knowledge reasoning module 14, wherein:

[0150] Detection module 11 is used to detect abnormal events based on multi-source heterogeneous data corresponding to key areas of the power grid, and to determine the event data packet corresponding to at least one key triggering event occurring in the key area.

[0151] The causal reasoning module 12 is used to input event data packets into a few-shot causal learning model based on a meta-learning framework for causal reasoning, and output causal datasets of each key triggering event; the few-shot causal learning model is used to learn causal association information between different power grid events, including events that do not occur frequently in the power grid;

[0152] The association processing module 13 is used to perform association processing on the causal dataset of each key triggering event and the current dynamic causal knowledge graph of the power grid to obtain the causal chain of each key triggering event.

[0153] The knowledge reasoning module 14 is used to perform knowledge reasoning based on each causal chain and determine the early warning report for each key triggering event.

[0154] The power grid abnormal event early warning device provided in this application embodiment can be used to execute the technical solution in the above-described power grid abnormal event early warning method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0155] In one embodiment, the detection module 11 includes: an event detection unit and a data packet construction unit, wherein:

[0156] The event detection unit is used to detect events based on multi-source heterogeneous data and obtain structured event information corresponding to at least one abnormal related event occurring in the key area.

[0157] The data packet construction unit is used to construct event data packets corresponding to each key triggering event based on the structured event information and multi-source heterogeneous data of each abnormal associated event when there is a key triggering event in each abnormal associated event.

[0158] The power grid abnormal event early warning device provided in this application embodiment can be used to execute the technical solution in the above-described power grid abnormal event early warning method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0159] In one embodiment, the event detection unit is specifically used for:

[0160] Time alignment is performed on multi-source heterogeneous data to obtain aligned heterogeneous data;

[0161] The aligned heterogeneous data is input into a cross-modal attention network model for power grid event detection, and the power grid event identifier of at least one power grid event occurring in the key area is output.

[0162] Based on a lightweight deep learning model, target detection is performed according to the power grid event identifier of each power grid event and the heterogeneous event data corresponding to each power grid event, so as to obtain the structured event information corresponding to the abnormal related events in each power grid event.

[0163] The power grid abnormal event early warning device provided in this application embodiment can be used to execute the technical solution in the above-described power grid abnormal event early warning method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0164] In one embodiment, the power grid anomaly early warning device further includes: a map update module, wherein:

[0165] The graph update module is used to update the initial causal knowledge graph of the power grid based on the causal dataset of each key triggering event, so as to obtain a dynamic causal knowledge graph.

[0166] The power grid abnormal event early warning device provided in this application embodiment can be used to execute the technical solution in the above-described power grid abnormal event early warning method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0167] In one embodiment, the association processing module 13 includes: a semantic matching unit and an information chain extraction unit, wherein:

[0168] The semantic matching unit is used to perform semantic matching on any key triggering event based on the causal dataset of the key triggering event and the current dynamic causal knowledge graph of the power grid, and obtain the matching node in the dynamic causal knowledge graph.

[0169] The information chain extraction unit is used to extract risk behavior chains associated with matching nodes from the dynamic causal knowledge graph as causal chains of key triggering events.

[0170] The power grid abnormal event early warning device provided in this application embodiment can be used to execute the technical solution in the above-described power grid abnormal event early warning method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0171] In one embodiment, the knowledge reasoning module 14 includes: a risk assessment unit and a reasoning explanation unit, wherein:

[0172] The risk assessment unit is used to assess the risk level of any critical triggering event based on the causal chain of the critical triggering event, the preset risk severity of the terminal nodes in the causal chain, the corresponding link confidence and link complexity.

[0173] The reasoning and interpretation unit is used to input the causal chain and the corresponding risk level into a pre-built large language model for enhanced reasoning and language interpretation, and to obtain early warning reports corresponding to key triggering events.

[0174] The power grid abnormal event early warning device provided in this application embodiment can be used to execute the technical solution in the above-described power grid abnormal event early warning method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0175] In one embodiment, the above-mentioned power grid anomaly early warning device further includes: a report sending module, wherein:

[0176] The report sending module is used to send early warning reports of each key triggering event to the user terminal, instructing the user to take security measures for each key triggering event.

[0177] The power grid abnormal event early warning device provided in this application embodiment can be used to execute the technical solution in the above-described power grid abnormal event early warning method embodiment of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0178] Each module in the aforementioned power grid anomaly early warning device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0179] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 7 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores historical and current power grid operation data. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a method for early warning of abnormal power grid events.

[0180] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0181] In one embodiment, a computer device is also provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the technical solution in the above-described power grid abnormal event early warning method embodiment of this application. The implementation principle and technical effect are similar, and will not be repeated here.

[0182] In one embodiment, a computer-readable storage medium is also provided, on which a computer program is stored. When the computer program is executed by a processor, it implements the technical solution of the above-described power grid abnormal event early warning method of this application. Its implementation principle and technical effect are similar, and will not be repeated here.

[0183] In one embodiment, a computer program product is also provided, including a computer program that, when executed by a processor, implements the technical solution of the above-described power grid abnormal event early warning method of this application. Its implementation principle and technical effect are similar and will not be repeated here.

[0184] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one of relational databases and non-relational databases. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0185] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0186] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for early warning of abnormal events in a power grid, characterized in that, The method includes: Time alignment is performed on the multi-source heterogeneous data corresponding to the key area of ​​the power grid to obtain aligned heterogeneous data. The aligned heterogeneous data is then input into a cross-modal attention network model for power grid event detection, and the power grid event identifier of at least one power grid event occurring in the key area is output. Based on a lightweight deep learning model, target detection is performed according to the power grid event identifier of each power grid event and the event heterogeneous data corresponding to each power grid event, so as to obtain the structured event information corresponding to the abnormal related events in each power grid event; In the case where a key triggering event exists in each of the aforementioned abnormal associated events, an event data packet corresponding to at least one key triggering event occurring in the key area is constructed based on the structured event information of each of the aforementioned abnormal associated events and the multi-source heterogeneous data. The event data packets are input into a few-shot causal learning model based on a meta-learning framework for causal inference, and the causal datasets of each key triggering event are output. The few-shot causal learning model is used to learn the causal association information between different power grid events, including events that do not occur frequently in the power grid. The causal datasets of each key triggering event and the current dynamic causal knowledge graph of the power grid are correlated to obtain the causal chain of each key triggering event. The dynamic causal knowledge graph adopts a three-layer node system of entity-event-risk and a two-layer edge relationship system of causality-spatiotemporal relationship. The nodes include event nodes, equipment nodes, environment nodes and parameter nodes, and the edges include causal edges that represent the induction or cause of the relationship and association edges that represent physical connection or attribute association. Each causal chain and its corresponding risk level are input into a pre-built large language model for enhanced reasoning and language interpretation to obtain early warning reports for each key triggering event. The early warning reports are in natural language text and include at least a risk description, cause tracing, and specific handling suggestions.

2. The method according to claim 1, characterized in that, The process of acquiring the dynamic causal knowledge graph includes: Based on the causal datasets of each of the key triggering events, the initial causal knowledge graph of the power grid is updated to obtain the dynamic causal knowledge graph.

3. The method according to claim 1, characterized in that, The process of associating the causal datasets of each of the key triggering events with the current dynamic causal knowledge graph of the power grid to obtain the causal chain of each of the key triggering events includes: For any key triggering event, semantic matching is performed based on the causal dataset of the key triggering event and the current dynamic causal knowledge graph of the power grid to obtain the matching node in the dynamic causal knowledge graph; The risk behavior chain associated with the matching node is extracted from the dynamic causal knowledge graph as the causal chain of the key triggering event.

4. The method according to any one of claims 1-3, characterized in that, The method further includes: For any critical triggering event, a risk assessment is performed based on the causal chain of the critical triggering event, the preset risk severity of the end node in the causal chain, the corresponding link confidence and link complexity, to obtain the risk level of the causal chain.

5. The method according to any one of claims 1-3, characterized in that, The method further includes: The warning reports for each of the aforementioned key triggering events are sent to the user terminal, instructing that each of the aforementioned key triggering events be handled securely.

6. A power grid anomaly early warning device, characterized in that, The device includes: The detection module is used to detect abnormal events based on multi-source heterogeneous data corresponding to key areas of the power grid, and to determine the event data packet corresponding to at least one key triggering event occurring in the key area. The causal reasoning module is used to input the event data packet into a few-shot causal learning model based on a meta-learning framework for causal reasoning, and output causal datasets for each of the key triggering events; the few-shot causal learning model is used to learn causal association information between different power grid events, and the power grid events include events that do not occur frequently in the power grid; The association processing module is used to perform association processing on the causal datasets of each key triggering event and the current dynamic causal knowledge graph of the power grid to obtain the causal chain of each key triggering event. The dynamic causal knowledge graph adopts a three-layer node system of entity-event-risk and a two-layer edge relationship system of causality-spatiotemporal relationship. The nodes include event nodes, equipment nodes, environment nodes and parameter nodes, and the edges include causal edges that represent the induction or cause of the relationship and association edges that represent physical connection or attribute association. The knowledge reasoning module is used to input each of the causal chains and the corresponding risk levels into a pre-built large language model for enhanced reasoning and language interpretation, and to obtain early warning reports for each of the key triggering events; the early warning reports are natural language text, and the early warning reports include at least a risk description, cause tracing and specific handling suggestions; The detection module is specifically used for: The multi-source heterogeneous data is time-aligned to obtain aligned heterogeneous data. The aligned heterogeneous data is then input into a cross-modal attention network model for power grid event detection, and the power grid event identifier of at least one power grid event occurring in the key area is output. Based on a lightweight deep learning model, target detection is performed according to the power grid event identifier of each power grid event and the event heterogeneous data corresponding to each power grid event, so as to obtain the structured event information corresponding to the abnormal related events in each power grid event; If a key triggering event exists in each of the aforementioned abnormal associated events, an event data packet corresponding to at least one key triggering event occurring within the key area is constructed based on the structured event information of each of the aforementioned abnormal associated events and the multi-source heterogeneous data.

7. The apparatus according to claim 6, characterized in that, The power grid abnormal event early warning device also includes: The graph update module is used to update the initial causal knowledge graph of the power grid based on the causal dataset of each of the key triggering events, so as to obtain the dynamic causal knowledge graph.

8. The apparatus according to claim 6, characterized in that, The association processing module includes: A semantic matching unit is used to perform semantic matching on any key triggering event based on the causal dataset of the key triggering event and the current dynamic causal knowledge graph of the power grid, so as to obtain the matching node in the dynamic causal knowledge graph. The information chain extraction unit is used to extract the risk behavior chain associated with the matching node from the dynamic causal knowledge graph as the causal chain of the key triggering event.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images