Electric power system fault tracing and early warning method and system based on AI large model

Abstract

The invention relates to the technical field of fault diagnosis, in particular to a power system fault tracing and early warning method and system based on an AI large model, and the method comprises the steps: collecting multi-source heterogeneous data from a power system, carrying out the preprocessing, deep feature extraction and enhancement, mapping a multi-modal feature tensor in a feature subset to a unified feature space, and carrying out the recognition of the multi-modal feature tensor. Identifying the mapped multi-modal feature tensor by using a pre-trained AI large model to obtain a state feature representation and a risk score; performing similarity calculation on the state feature representation and a current state reference vector, and when an abnormal state is determined, constructing a fault propagation graph for a fault feature vector by using a graph neural network, and marking a fault source and an influence path at the same time; and the state corresponding to the next sliding time window is subjected to rolling prediction based on the set time step length, and when the risk score exceeds the set dynamic threshold value, multi-level early warning is triggered, so that the fault tracing accuracy and the early warning efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and in particular to a method and system for tracing and early warning of power system faults based on a large AI model. Background Technology

[0002] The power system is a crucial component of modern society's critical infrastructure, and its safe and stable operation directly impacts the national economy and people's lives. Traditional power system fault management primarily relies on human experience, physical model-based simulation tools, or simple rule engines. For example, when a fault occurs, maintenance personnel analyze monitoring data (such as voltage, current, and power) and event records, combining historical experience to locate the fault and analyze its causes. Meanwhile, early warning mechanisms are mostly based on threshold alarms or statistical models, such as using traditional machine learning methods like support vector machines and decision trees to detect anomalies in real-time data. However, with the expansion of power system scale and the increasing complexity of its structure (e.g., the integration of new energy sources and the development of smart grids), the volume of data has surged and become multi-source and heterogeneous (including SCADA data, PMU data, equipment status data, and meteorological data), posing significant challenges to traditional methods. These methods often have limited processing capabilities, making it difficult to extract deep features from massive amounts of data, resulting in low accuracy in fault tracing and high early warning delays. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for fault tracing and early warning in power systems based on AI large models, so as to improve the accuracy of fault tracing and the efficiency of early warning.

[0004] To achieve the above objectives, in a first aspect, the present invention provides a method for power system fault tracing and early warning based on an AI large model, comprising the following steps:

[0005] Multi-source heterogeneous data is collected from the power system, and after preprocessing the multi-source heterogeneous data, deep feature extraction and enhancement are performed to obtain a feature subset;

[0006] The multimodal feature tensors in the feature subset are mapped to a unified feature space, and the pre-trained AI large model is used to identify the mapped multimodal feature tensors to obtain state feature representation and risk score.

[0007] The similarity between the state feature representation and the current state baseline vector is calculated. When an abnormal state is determined, a fault propagation graph is constructed from the fault feature vector using a graph neural network, and the root cause and impact path of the fault are labeled.

[0008] Based on a set time step, the state corresponding to the next sliding time window is predicted in a rolling manner, and when the risk score exceeds the set dynamic threshold, a multi-level warning is triggered.

[0009] This involves collecting multi-source heterogeneous data from the power system, preprocessing the multi-source heterogeneous data, and then performing deep feature extraction and enhancement to obtain a feature subset, including:

[0010] Data is collected from multiple dimensions in the power system, including power generation, transmission, transformation and distribution. The collected data packages include real-time operation monitoring data, equipment status monitoring data, environmental factor data and historical data.

[0011] The collected multi-source heterogeneous data is cleaned and repaired, and then standardized after data alignment.

[0012] For the standardized data, deep features are extracted based on the time domain and frequency domain respectively, and feature enhancement is performed in combination with physical rules to obtain feature subsets.

[0013] Specifically, the multimodal feature tensors in the feature subset are mapped to a unified feature space, and a pre-trained AI large model is used to identify the mapped multimodal feature tensors to obtain state feature representations and risk scores, including:

[0014] The multimodal feature tensors in the feature subset are mapped to a unified feature space through a neural network model constructed from multilayer perceptrons.

[0015] The multimodal feature tensors in the unified feature space are trained and identified using a pre-trained AI large model to obtain state feature representations and risk scores.

[0016] The pre-training process for large AI models includes:

[0017] The acquired normal historical data is masked from three dimensions: time, space, and features. The masked feature blocks are then predicted using a large AI model.

[0018] Data samples from two time windows are extracted from the normal historical data, processed separately, and used as positive and negative samples for training.

[0019] The method further includes:

[0020] The output head of the pre-trained AI model is replaced with a classifier and a regressor, and a historical sample memory is constructed. The historical sample memory is used to call up old data for training when the pre-trained AI model is training on the currently collected multi-source heterogeneous data.

[0021] Specifically, the similarity between the state feature representation and the current state baseline vector is calculated, and when an abnormal state is determined, a fault propagation graph is constructed from the fault feature vector using a graph neural network, while simultaneously labeling the fault root cause and impact path, including:

[0022] The encoder in the pre-trained AI large model is used to convert the state feature representation into a state representation vector. The similarity between the state representation vector and the current state baseline vector is calculated. When an abnormal state is determined, the state representation vector is marked as a fault feature vector.

[0023] A fault propagation graph is constructed from fault feature vectors using a graph neural network, while simultaneously labeling the fault root cause and the impact path.

[0024] Specifically, a fault propagation graph is constructed from fault feature vectors using a graph neural network, simultaneously labeling the fault root cause and its impact path, including:

[0025] A basic knowledge graph is constructed based on the static topology of the power grid, where nodes represent power equipment and edges represent electrical connections.

[0026] The fault feature vector generated by the AI ​​model is decomposed and mapped to the corresponding device nodes. The configuration logic of the protection device and the physical propagation law of the fault are embedded into the graph in the form of rules to obtain the fault propagation graph, while marking the root cause of the fault and the impact path.

[0027] After labeling the root cause of the fault and the impact path, the method further includes:

[0028] The contribution value of the multimodal feature tensor of the generated fault feature vector is calculated based on the SHAP method. Combined with the self-attention weight results of the AI ​​large model and the fault propagation graph, a diagnostic report is generated.

[0029] Secondly, the present invention provides a power system fault tracing and early warning system based on an AI large-scale model, which is applied to the power system fault tracing and early warning method based on an AI large-scale model provided in the first aspect. The power system fault tracing and early warning system based on an AI large-scale model includes a data acquisition and processing module, a model training and identification module, a fault tracing module, and an early warning module.

[0030] The data acquisition and processing module is used to acquire multi-source heterogeneous data from the power system, and after preprocessing the multi-source heterogeneous data, perform deep feature extraction and enhancement to obtain a feature subset.

[0031] The model training and recognition module is used to map the multimodal feature tensors in the feature subset to a unified feature space, and use the pre-trained AI large model to recognize the mapped multimodal feature tensors to obtain state feature representation and risk score.

[0032] The fault tracing module is used to calculate the similarity between the state feature representation and the current state baseline vector, and when it is determined to be an abnormal state, it uses a graph neural network to construct a fault propagation graph for the fault feature vector, while marking the fault root cause and the impact path.

[0033] The early warning module is used to perform rolling prediction of the state corresponding to the next sliding time window based on a set time step, and to trigger a multi-level early warning when the risk score exceeds a set dynamic threshold.

[0034] This invention discloses a power system fault tracing and early warning method and system based on an AI large-scale model. The AI ​​large-scale model-based power system fault tracing and early warning system includes a data acquisition and processing module, a model training and identification module, a fault tracing module, and an early warning module. It collects multi-source heterogeneous data from the power system, preprocesses the multi-source heterogeneous data, performs deep feature extraction and enhancement to obtain a feature subset; maps the multimodal feature tensors in the feature subset to a unified feature space, and uses a pre-trained AI large-scale model to identify the mapped multimodal feature tensors, obtaining state feature representations and risk scores; calculates the similarity between the state feature representations and the current state baseline vector, and when an abnormal state is determined, uses a graph neural network to construct a fault propagation graph from the fault feature vectors, simultaneously labeling the fault root cause and impact path; performs rolling prediction of the state corresponding to the next sliding time window based on a set time step, and triggers multi-level early warnings when the risk score exceeds a set dynamic threshold, thereby improving the accuracy of fault tracing and the efficiency of early warning. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0036] Figure 1 This is a schematic diagram illustrating the steps of a power system fault tracing and early warning method based on an AI large model, according to the first embodiment of the present invention.

[0037] Figure 2 This is a flowchart illustrating a power system fault tracing and early warning method based on an AI large model provided by the present invention.

[0038] Figure 3 This is a schematic diagram of the structure of a power system fault tracing and early warning system based on an AI large model, according to the second embodiment of the present invention.

[0039] Figure 4 This is a schematic diagram of the electronic device of the present invention.

[0040] In the diagram: 101 - Data acquisition and processing module, 102 - Model training and identification module, 103 - Fault tracing module, 104 - Early warning module. Detailed Implementation

[0041] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0042] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0043] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0044] The first embodiment of this application is as follows:

[0045] Please see Figures 1-2 This invention provides a method for power system fault tracing and early warning based on an AI large model, comprising the following steps:

[0046] S101. Collect multi-source heterogeneous data from the power system, and after preprocessing the multi-source heterogeneous data, perform deep feature extraction and enhancement to obtain a feature subset.

[0047] Specifically, the data acquisition stage is responsible for acquiring multi-dimensional operational data from various components of the power system in real-time or near real-time. This data is collected through various sensors, smart terminals, and protection devices deployed in power generation, transmission, substation, and distribution stages. The collected data is mainly divided into the following categories:

[0048] Real-time operational monitoring data primarily originates from data acquisition and monitoring systems, including remote measurements of voltage, current, active power, reactive power, and frequency at key power grid nodes. Synchronous phasor measurement device data provides higher-precision synchronous phasor information, which is crucial for analyzing the dynamic behavior of the power grid.

[0049] Equipment condition monitoring data comes from online monitoring devices installed on critical equipment, such as oil chromatography data, winding temperature, and bushing dielectric loss values ​​for transformers, as well as opening and closing coil currents and energy storage status for circuit breakers. This data directly reflects the health status of the power equipment itself.

[0050] Environmental and external factor data include meteorological data such as temperature, humidity, wind speed, precipitation, and lightning location information, as well as geographic information data such as the topography of the line corridor. These external factors are important conditions that induce or affect faults.

[0051] Historical events and knowledge data: This includes historical fault records, protection action information, switch change sequence, equipment ledgers, power grid topology diagrams, etc. This data constitutes the background knowledge and historical basis for fault analysis.

[0052] All the raw data collected is transmitted to the data acquisition and processing module 101 for centralized processing via a secure and reliable data bus or communication network.

[0053] Raw data inevitably contains noise, outliers, and missing values ​​during acquisition and transmission, which can negatively impact model performance if used directly. For outliers, in addition to traditional statistical methods such as interquartile range (ICM) thresholding, we introduce preliminary anomaly detection based on pre-trained features from a large AI model. The model performs an initial scan of the data stream, identifying data points that significantly deviate from normal operating conditions, and then performs secondary verification based on equipment physical constraints, effectively distinguishing between genuine fault signals and acquisition interference.

[0054] To address the issue of missing data, a collaborative repair strategy based on spatiotemporal correlation is employed. This method not only considers the temporal continuity of a single measuring point but also fully utilizes the topological correlation of the power system. For example, if current data for a certain line is missing, the most reasonable interpolation value is generated by referring to the data of its upstream and downstream nodes and the operating status of parallel lines in the same corridor, using a graph neural network instead of simple linear interpolation, greatly improving the physical rationality of the repaired data.

[0055] Because the timestamp precision, acquisition frequency, and transmission delay of multi-source data vary, direct integration can lead to timing discrepancies. Therefore, a high-precision clock source is used as a reference to assign a unified timestamp to all incoming data. For data of different frequencies, an event-triggered dynamic alignment strategy is employed. For example, using high-frequency synchronous phasor measurement device data as a reference, when a specific event such as a voltage drop occurs, lower-frequency data acquisition and monitoring system data, equipment status data, etc., before and after that point in time are automatically aligned and extracted to form a multimodal data slice within a unified time window centered on the event.

[0056] After aligning the multi-source data, a unified multi-dimensional feature vector is constructed. The data within each smallest analysis time unit is organized into a structured data volume, which not only contains numerical information but also includes metadata tags such as data source and physical meaning, laying a solid foundation for subsequent feature extraction.

[0057] To eliminate the influence of different physical dimensions and numerical ranges on the training of large AI models, data normalization is necessary. A hierarchical normalization strategy is adopted, selecting the most suitable normalization method based on the characteristics of different types of data.

[0058] For operational monitoring data and other data conforming to a Gaussian distribution, Z-Score standardization is used to make its mean zero and standard deviation one. For equipment status data such as temperature with clear upper and lower limits, max-min normalization is used to scale it to the range of zero to one. The normalization parameters are not fixed, but are dynamically adjusted according to different system operating conditions, such as distinguishing between summer peak load and winter normal load, so that the normalized data can better reflect relative changes and improve the model's adaptability to different operating modes.

[0059] Directly inputting simple raw data into the model is inefficient. For the standardized data, deep feature extraction is performed in both the time and frequency domains, and feature enhancement is combined with physical rules to obtain a feature subset, specifically:

[0060] Time-domain and statistical feature extraction: From the normalized time-series data, a sliding window is used to extract statistical features such as mean, variance, skewness, and kurtosis within each window to characterize the static distribution characteristics of the data. Simultaneously, waveform indicators such as peak factor and margin factor are extracted to reflect the impulse characteristics of the signal.

[0061] Frequency Domain and Time-Series Feature Extraction: Using signal processing methods such as Fast Fourier Transform and Wavelet Transform, the time-series signal is converted to the frequency domain to extract features such as energy distribution and spectral centroid in specific frequency bands, effectively capturing frequency components related to mechanical faults and resonances. Simultaneously, the autocorrelation coefficient and partial autocorrelation coefficient of the time-series data are calculated to reveal its dynamic evolution.

[0062] By combining physics knowledge with data-driven approaches, for example, Kirchhoff's laws are used to automatically calculate whether the power at critical nodes is balanced, and the imbalance is used as an important derived feature. Based on the equipment's thermal model, the theoretical temperature is calculated from the ambient temperature and load current, and the deviation from the measured temperature is used as a characteristic of equipment anomalies. This feature enhancement based on physics rules ensures that the data input into the AI ​​model not only exhibits statistical regularities but also contains profound insights into the operating mechanisms of the power system, greatly improving the interpretability and effectiveness of feature engineering.

[0063] Feature selection techniques are used to evaluate the relevance of each feature to the target, eliminate redundant features, and form an optimal feature subset. Ultimately, each analyzed sample is represented as a unified multimodal feature tensor.

[0064] This multimodal feature tensor is the final output of the preprocessing stage. It is a structured, high-quality, and information-rich numerical matrix that perfectly integrates multi-source heterogeneous information from the power system. This tensor will serve as direct input to large AI models, providing a solid data foundation for accurate fault tracing and reliable early warning.

[0065] S102. Map the multimodal feature tensors in the feature subset to a unified feature space, and use a pre-trained AI large model to identify the mapped multimodal feature tensors to obtain state feature representation and risk score.

[0066] Specifically, the multimodal feature tensor obtained after feature extraction includes various types of features such as time domain, frequency domain, statistical, and physical knowledge-guided features. Although these features have been standardized, their distribution in mathematical space and their inherent semantics still differ. Directly concatenating them and inputting them into the model makes it difficult for the model to fully understand the deep correlations between different feature modalities.

[0067] Introducing a learnable feature mapping network, a lightweight neural network typically composed of multilayer perceptrons, is tasked with acting as an "intelligent translator," nonlinearly mapping features from different sources and with different physical meanings to a novel, high-dimensional, unified feature space.

[0068] In this unified feature space, the goal of the mapping is to make:

[0069] Features with similar semantics are close: For example, the temperature feature representing "overload" and the operation monitoring feature representing "excessive current" will have very similar vector representations in a unified space, even if the original values ​​and dimensions are different.

[0070] Separation of semantically distinct features: For example, the voltage drop caused by lightning strikes and the slow failure caused by equipment insulation aging are clearly distinguished in a unified space.

[0071] Preserving spatiotemporal correlation: This mapping process takes into account the spatiotemporal context of features. For example, similar anomaly features at different monitoring points on the same line will reflect their topological proximity in the mapping results in a unified space.

[0072] The formation of this unified feature space is not based on manual rules, but rather on self-supervised learning through subsequent pre-training tasks. Ultimately, all heterogeneous features are transformed into a unified, semantically rich, high-dimensional sequence of numerical vectors, which is what is truly input into the core architecture of large AI models (such as Transformers).

[0073] Then, the pre-trained AI large model is used to train and identify the multimodal feature tensor in the unified feature space to obtain state feature representation and risk score.

[0074] The goal of pre-training large AI models is to enable them to learn the inherent "grammar" and "language patterns"—that is, the spatiotemporal dynamics of the power system under normal operation—from massive amounts of unlabeled historical data. This is a form of self-supervised learning that requires no manual labeling and greatly expands the scale of data available for training.

[0075] The pre-training process involves two self-supervised processes. The first is a multi-level masking reconstruction task, specifically where the model randomly masks (covers) a subset of feature blocks from the uniform feature sequence of the input. These masking strategies are hierarchical.

[0076] Temporal masking: Randomly masking all features at a given point in time, allowing the model to reconstruct it based on data from previous and subsequent times. This forces the model to learn the temporal evolution of the power system.

[0077] Spatial layer masking: Randomly masking the features of all measurement points in a specific geographical location (such as a substation) allows the model to reconstruct the network based on data from neighboring sites in the topology. This forces the model to learn the topological connectivity and power balance relationships of the power grid.

[0078] Feature layer masking: Randomly masking a certain feature of all measurement points (e.g., masking only all voltage values) allows the model to reconstruct based on other coexisting features (e.g., current, power). This forces the model to learn the coupling relationships between different physical quantities.

[0079] The core of the model is the Transformer encoder, which, through its self-attention mechanism, can simultaneously focus on contextual information in the spatiotemporal dimensions and strive to predict the masked feature blocks. Through this task, the model learns deep, dynamic spatiotemporal dependency patterns in power system data.

[0080] The second approach is temporal contrastive learning, primarily aimed at enhancing the model's sensitivity to anomalies and subtle changes. Contrastive learning is introduced by extracting data samples from two time windows from normal data: one serves as an anchor point, and the other is modified by adding slight noise or time scaling to serve as a positive sample. Samples are then extracted from data from different times or different pre-fault indicators as negative samples. The model's goal is to narrow the distance between the anchor point and positive samples in a unified feature space, while simultaneously widening the distance between the anchor point and negative samples. This task enables the model to learn a "metric," namely the ability to determine whether two states are similar, which is crucial for subsequent detection of anomalies deviating from normal patterns.

[0081] The combined effect of these two pre-training tasks transforms the large AI model from a general language model into a domain model that deeply understands the "physical language" of the power system, laying a solid foundation for accurate judgment in the next step.

[0082] The pre-trained model possesses rich domain knowledge, but it has not yet been specifically optimized for the two tasks of "fault tracing" and "early warning". The model should be precisely tuned using historical data with accurate labels (such as fault records, fault types, and root cause devices).

[0083] A multi-task learning framework was designed. The model shares the same feature extraction backbone network during fine-tuning, but has two output heads: a fault tracing output head, which is a classifier responsible for predicting the type of fault and the most likely root cause device; and a risk warning output head, which is a regressor responsible for outputting a continuous risk score, representing the probability that the system tends to fail.

[0084] Through joint training, the features learned by the model can be used for both classification and regression, making the feature representation more robust and discriminative. Risk warning tasks can learn from data that is in poor condition but has not yet failed, while fault tracing tasks learn from failures that have already occurred; the two complement each other's knowledge.

[0085] Because power systems are dynamic, the commissioning of new equipment and changes in operating modes can cause data distribution drift. Traditional periodic full retraining is costly. Continuous monitoring of model performance automatically triggers a fine-tuning process when a decrease in prediction confidence or persistent prediction bias is detected. To avoid "catastrophic forgetting" (i.e., the model forgetting old knowledge while learning new knowledge), a memory replay technique is introduced. A core memory bank storing representative historical samples is built and maintained. During a new round of online fine-tuning, not only new data is used, but a portion of old data is also sampled from the core memory bank for training. This ensures that the model adapts to new changes without losing previously mastered important patterns.

[0086] Furthermore, the warning threshold is not a fixed value, but is dynamically adjusted based on the model's assessment of the overall operating status of the current system. When the system load is heavy or the environment is harsh, the threshold is appropriately relaxed to reduce false alarms; when the system is operating smoothly, the threshold is tightened to improve monitoring sensitivity. This adjustment process itself can also be optimized through reinforcement learning.

[0087] S103. Calculate the similarity between the state feature representation and the current state reference vector, and when it is determined to be an abnormal state, use a graph neural network to construct a fault propagation graph for the fault feature vector, and mark the fault root cause and the impact path.

[0088] Specifically, fault tracing analysis is triggered by the system's real-time anomaly detection mechanism. The core of the AI ​​large model—the Transformer encoder—processes the continuously input, mapped, unified feature sequence, outputting a high-dimensional state representation vector for each smallest analysis time unit. This vector condenses the overall operational state information of the current system from the perspective of multi-source data.

[0089] A "health status baseline vector set" representing the system's historical normal operation is constructed and maintained, generated by a large AI model. This baseline set is not static but dynamically evolves with typical operating conditions such as seasons and load levels. The real-time generated status representation vectors are compared with the corresponding health status baseline vectors under the current operating conditions, using a combination of cosine similarity and Euclidean distance. When the similarity falls below a dynamic threshold obtained through statistical learning, the system determines the current state as abnormal and triggers the fault tracing analysis process. Once the anomaly is confirmed, the status representation vectors within the time window before and after the triggering time are specially marked as key fault feature vectors. These key fault feature vectors represent a deep abstraction and essence extraction of the system's multimodal information during the fault period by the large AI model, and they will become the core input for subsequent tracing analysis.

[0090] To understand how faults propagate and evolve within the power grid topology, a basic knowledge graph is first constructed based on the grid's static topology (such as line connections, circuit breaker locations, and transformer relationships). Nodes in the graph represent electrical equipment, and edges represent electrical connections. This graph is dynamically enhanced; it not only contains the static topology but also injects two types of dynamic information in real time:

[0091] 1. Real-time operational status injection: The key fault feature vectors generated by the AI ​​large model are decomposed and mapped to the relevant device nodes. Each device node is no longer an abstract symbol, but is given a vector attribute from a unified feature space that represents its current health status.

[0092] 2. Protection Logic and Causal Rule Injection: The configuration logic of the protection device (such as overcurrent protection settings and operating time limits) and common fault physical propagation laws (such as the direction of short-circuit current) are embedded into the graph in the form of rules as constraints or propagation probabilities on the edges.

[0093] Graph neural networks (Graph Neural Networks) are designed to process graph-structured data with rich node attributes and edge relationships. Key fault feature vectors are used as initial node feature inputs to the Graph Neural Network. The inference process of the Graph Neural Network is an iterative message-passing process, simulating fault propagation.

[0094] Message passing: Each device node receives information from its neighboring nodes (based on the topology). The received information includes not only the state feature vectors of the neighboring nodes, but also the physical rules on the connecting edges, such as "downstream failure may cause upstream overcurrent".

[0095] Node state update: Each node aggregates information from its neighbors and combines it with its own initial critical fault feature vector. Through a learnable update function, it calculates a new node state representation that includes neighbor context information. This process iterates multiple times to simulate the gradual spread of fault impacts across multi-hop devices.

[0096] Root cause probability calculation: After multiple iterations, each node obtains a stable final state representation. A specific output layer of the graph neural network calculates the probability that each node is the root cause of the failure based on this final representation. The higher the initial anomalousness of a node, and the better its location can explain the anomalousness of other nodes through the topological path, the higher its root cause probability.

[0097] Ultimately, the graph neural network outputs a fault propagation graph. This graph, based on the original dynamic knowledge graph, clearly marks the root cause and impact path of the fault using visual elements (such as node color depth representing the degree of anomaly and arrow thickness representing the intensity of impact).

[0098] The most likely root cause of the failure (the device with the highest probability of being the root cause);

[0099] The main propagation path of the fault (the connecting edge with the greatest impact);

[0100] The range of affected equipment.

[0101] This also includes: utilizing interpretable AI technologies, such as SHAP-based methods, to retrospectively analyze which original features in the multimodal feature tensors initially input into the large AI model contributed most to the final generated key fault feature vector. For example, the analysis results show that "the third harmonic content of the current in line A" and "the dissolved hydrogen concentration in the oil of associated transformer B" are key factors triggering anomaly judgments, providing maintenance personnel with direct, data-driven evidence. Analyzing the weights of the self-attention mechanism within the large AI model reveals which time points and spatial locations (which substations or lines) the model focuses on when judging faults. This provides an explanation from the perspective of the spatiotemporal evolution of faults. Automatically synthesizing the fault propagation graph output by the graph neural network, the feature contribution tracing results, and the cross-modal attention analysis conclusions, a structured fault diagnosis report is generated. This report not only points out "where it's broken," but also explains "why it's judged to be broken" and "how the fault affects the system," forming a logical closed loop and greatly improving the credibility and practicality of the tracing results.

[0102] S104. Based on the set time step, perform rolling prediction of the state corresponding to the next sliding time window, and trigger multi-level early warning when the risk score exceeds the set dynamic threshold.

[0103] Specifically, firstly, the large AI model receives a unified sequence of feature space vectors from the current and historical data. The temporal prediction output head within the model, dedicated to early warning, is essentially a sequence-to-sequence predictor that combines the advantages of temporal convolutional networks in extracting local dependencies with the Transformer attention mechanism in capturing long-term associations.

[0104] At the start of the prediction process, the model predicts a system state representation vector for a specific future time step based on the latest data window. Then, using this prediction as part of the known input, the time window slides to predict the state for the next time step, and this process is repeated to form a multi-step prediction. This process is called "rolling prediction," and it generates a trajectory of the system's state over a short period of time.

[0105] Crucially, the model's predicted output is not the original voltage and current values, but rather a unified feature space vector at each future time point. This means the prediction directly reflects the system's future "health status" encoding. The system quantifies future risk by calculating the deviation of this future state trajectory from the "health status baseline vector set," generating a continuous, time-varying dynamic risk scoring curve. This curve demonstrates the speed and intensity of risk evolution over time, providing a precise basis for tiered early warning.

[0106] There are three levels of early warning: yellow, orange, and red.

[0107] Yellow Alert (Attention Level): Indicates a slight deviation in the system, suggesting potential risks.

[0108] Orange alert (alert level): Indicates that the system anomaly is worsening and the probability of failure is significantly increased.

[0109] Red Alert (Emergency Level): Indicates that the system is in a critical state, and a failure may be imminent or has already occurred.

[0110] Each warning level corresponds to a risk score threshold, but these thresholds are not fixed. An independent reinforcement learning agent manages these thresholds. This agent aims to maximize the long-term benefits of system safety and stability by continuously learning through interaction with the environment. It considers factors such as current load levels, weather conditions, and equipment maintenance status. Its actions involve dynamically adjusting the warning thresholds for yellow, orange, and red levels.

[0111] For example, during typhoon season or peak load periods, the system's inherent risk level is higher. The reinforcement learning AI will appropriately raise the thresholds for orange and red alerts to avoid generating a large number of invalid alarms due to oversensitivity. Conversely, at night when the weather is good and the load is stable, the thresholds will be lowered to increase monitoring sensitivity and not miss any weak abnormal signals. This dynamic adjustment ensures the accuracy and effectiveness of the early warning system.

[0112] When the dynamic risk score curve predicted by the rolling forecast exceeds a certain dynamic threshold, the corresponding level of warning is immediately triggered.

[0113] When an alert is triggered, the system automatically constructs a complete "alert context." This context includes:

[0114] Predictive information: dynamic risk score that triggers an early warning, the slope of risk escalation, and the future risk trajectory.

[0115] Real-time source tracing information: Associates the key fault feature vectors output by the AI ​​large model fault tracing engine at the current moment, as well as the suspicious root cause devices and high-risk propagation paths initially identified by the graph neural network.

[0116] Operating conditions: current power grid topology, load distribution, weather conditions, etc.

[0117] Then, the current warning context is matched with a stored database of historical warnings and fault handling cases for similarity. The most similar historical cases are identified, and their proven effective handling solutions are referenced. Simultaneously, on the power grid knowledge graph, starting with suspected root cause devices and considering the warning level (representing the breadth and depth of impact), the consequences of various handling measures (such as disconnecting a line, adjusting generator output, or activating capacitor reactors) are simulated and extrapolated. By traversing the graph, the effectiveness of each measure in is assessed, including whether it can effectively isolate risks and whether it will lead to excessive power supply losses, thus generating a cost-benefit analysis recommendation.

[0118] For yellow alerts, the recommended response is "monitoring and preparation," such as "paying close attention to changes in the current of line A and notifying inspection personnel to be on standby."

[0119] For orange alerts, the recommended response is to upgrade to "proactive intervention," such as "recommending remote adjustment of the reactive power compensation device at substation B to prepare for switching the operating mode of line C."

[0120] For a red alert, the recommended action would be "emergency control," such as "it is recommended to immediately trip circuit breaker D and initiate load shedding scheme E."

[0121] All these suggestions will be clearly pushed to the dispatcher through the human-computer interaction interface, along with the reasoning, such as "This suggestion is based on the successful handling experience of a similar case on X month Y day in 2023" or "Graph analysis shows that this operation can minimize the power outage area".

[0122] The system records the actual measures taken by the dispatcher and the subsequent system recovery. This data forms new cases stored in the knowledge base and serves as reward signals for the reinforcement learning agent, used to optimize the dynamic early warning threshold strategy. This creates a continuous improvement loop from early warning to decision-making and then learning from feedback, making the entire system increasingly intelligent and accurate.

[0123] The second embodiment of this application is as follows:

[0124] Please see Figure 3 This invention provides a power system fault tracing and early warning system based on an AI large-scale model, applied to a power system fault tracing and early warning method based on an AI large-scale model as provided in the first embodiment. The AI ​​large-scale model-based power system fault tracing and early warning system includes a data acquisition and processing module 101, a model training and identification module 102, a fault tracing module 103, and an early warning module 104.

[0125] The data acquisition and processing module 101 is used to acquire multi-source heterogeneous data from the power system, and after preprocessing the multi-source heterogeneous data, perform deep feature extraction and enhancement to obtain a feature subset.

[0126] The model training and recognition module 102 is used to map the multimodal feature tensors in the feature subset to a unified feature space, and use the pre-trained AI large model to recognize the mapped multimodal feature tensors to obtain state feature representation and risk score.

[0127] The fault tracing module 103 is used to calculate the similarity between the state feature representation and the current state reference vector, and when it is determined to be an abnormal state, it uses a graph neural network to construct a fault propagation graph for the fault feature vector, while marking the fault root cause and the impact path.

[0128] The early warning module 104 is used to perform rolling prediction of the state corresponding to the next sliding time window based on a set time step, and to trigger a multi-level early warning when the risk score exceeds a set dynamic threshold.

[0129] Regarding the system in the above embodiments, the specific ways in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0130] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0131] Accordingly, this application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; and when the one or more programs are executed by the one or more processors, causing the one or more processors to implement the above-described method for power system fault tracing and early warning based on an AI large model. Figure 4 The diagram shown is a hardware structure diagram of any device with data processing capabilities in a power system fault tracing and early warning system based on an AI large model, as provided in an embodiment of the present invention. (Except for...) Figure 4In addition to the processor, memory, and network interface shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0132] Accordingly, this application also provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, implement the aforementioned method for tracing and early warning of power system faults based on a large AI model. The computer-readable storage medium can be an internal storage unit of any data-processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data-processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data-processing device, and can also be used to temporarily store data that has been output or will be output.

[0133] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0134] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for fault tracing and early warning in power systems based on a large AI model, characterized in that, Includes the following steps: Multi-source heterogeneous data is collected from the power system, and after preprocessing the multi-source heterogeneous data, deep feature extraction and enhancement are performed to obtain a feature subset; The multimodal feature tensors in the feature subset are mapped to a unified feature space, and the pre-trained AI large model is used to identify the mapped multimodal feature tensors to obtain state feature representation and risk score. The similarity between the state feature representation and the current state baseline vector is calculated. When an abnormal state is determined, a fault propagation graph is constructed from the fault feature vector using a graph neural network, and the root cause and impact path of the fault are labeled. Based on a set time step, the state corresponding to the next sliding time window is predicted in a rolling manner, and when the risk score exceeds the set dynamic threshold, a multi-level warning is triggered.

2. The method for power system fault tracing and early warning based on AI large model as described in claim 1, characterized in that, Multi-source heterogeneous data is collected from the power system. After preprocessing the multi-source heterogeneous data, deep feature extraction and enhancement are performed to obtain a feature subset, including: Data is collected from multiple dimensions in the power system, including power generation, transmission, transformation and distribution. The collected data packages include real-time operation monitoring data, equipment status monitoring data, environmental factor data and historical data. The collected multi-source heterogeneous data is cleaned and repaired, and then standardized after data alignment. For the standardized data, deep features are extracted based on the time domain and frequency domain respectively, and feature enhancement is performed in combination with physical rules to obtain feature subsets.

3. The method for power system fault tracing and early warning based on AI large model as described in claim 1, characterized in that, The multimodal feature tensors in the aforementioned feature subset are mapped to a unified feature space. A pre-trained AI model is then used to identify the mapped multimodal feature tensors, yielding state feature representations and risk scores, including: The multimodal feature tensors in the feature subset are mapped to a unified feature space through a neural network model constructed from multilayer perceptrons. The multimodal feature tensors in the unified feature space are trained and identified using a pre-trained AI large model to obtain state feature representations and risk scores.

4. The method for power system fault tracing and early warning based on AI large model as described in claim 3, characterized in that, The pre-training process of large AI models includes: The acquired normal historical data is masked from three dimensions: time, space, and features. The masked feature blocks are then predicted using a large AI model. Data samples from two time windows are extracted from the normal historical data, processed separately, and used as positive and negative samples for training.

5. The method for power system fault tracing and early warning based on AI large model as described in claim 4, characterized in that, The method further includes: The output head of the pre-trained AI model is replaced with a classifier and a regressor, and a historical sample memory is constructed. The historical sample memory is used to call up old data for training when the pre-trained AI model is training on the currently collected multi-source heterogeneous data.

6. The method for power system fault tracing and early warning based on AI large model as described in claim 1, characterized in that, The similarity between the state feature representation and the current state baseline vector is calculated. When an abnormal state is determined, a fault propagation graph is constructed from the fault feature vector using a graph neural network, while simultaneously labeling the fault root cause and impact path, including: The encoder in the pre-trained AI large model is used to convert the state feature representation into a state representation vector. The similarity between the state representation vector and the current state baseline vector is calculated. When an abnormal state is determined, the state representation vector is marked as a fault feature vector. A fault propagation graph is constructed from fault feature vectors using a graph neural network, while simultaneously labeling the fault root cause and the impact path.

7. The method for power system fault tracing and early warning based on AI large model as described in claim 6, characterized in that, A fault propagation graph is constructed from fault feature vectors using a graph neural network, simultaneously labeling the fault root cause and its impact path, including: A basic knowledge graph is constructed based on the static topology of the power grid, where nodes represent power equipment and edges represent electrical connections. The fault feature vector generated by the AI ​​model is decomposed and mapped to the corresponding device nodes. The configuration logic of the protection device and the physical propagation law of the fault are embedded into the graph in the form of rules to obtain the fault propagation graph, while marking the root cause of the fault and the impact path.

8. The method for power system fault tracing and early warning based on AI large model as described in claim 7, characterized in that, After identifying the root cause and impact path of the fault, the method further includes: The contribution value of the multimodal feature tensor of the generated fault feature vector is calculated based on the SHAP method. Combined with the self-attention weight results of the AI ​​large model and the fault propagation graph, a diagnostic report is generated.

9. A power system fault tracing and early warning system based on an AI large-scale model, applied to the power system fault tracing and early warning method based on an AI large-scale model as described in claim 1, characterized in that, The AI-based large-scale model-based power system fault tracing and early warning system includes a data acquisition and processing module, a model training and identification module, a fault tracing module, and an early warning module. The data acquisition and processing module is used to acquire multi-source heterogeneous data from the power system, and after preprocessing the multi-source heterogeneous data, perform deep feature extraction and enhancement to obtain a feature subset. The model training and recognition module is used to map the multimodal feature tensors in the feature subset to a unified feature space, and use the pre-trained AI large model to recognize the mapped multimodal feature tensors to obtain state feature representation and risk score. The fault tracing module is used to calculate the similarity between the state feature representation and the current state baseline vector, and when it is determined to be an abnormal state, it uses a graph neural network to construct a fault propagation graph for the fault feature vector, while marking the fault root cause and the impact path. The early warning module is used to perform rolling prediction of the state corresponding to the next sliding time window based on a set time step, and to trigger a multi-level early warning when the risk score exceeds a set dynamic threshold.

Images