Power grid equipment dynamic fault probability prediction method and system based on physical information and time-space deep learning in extreme weather

Abstract

The invention discloses a power grid equipment dynamic fault probability prediction method and system based on physical information and time-space deep learning in extreme weather, and the method comprises the steps: obtaining multi-source heterogeneous data of power grid equipment in a target region, carrying out the preprocessing of the multi-source heterogeneous data, and obtaining a standard data set; constructing advanced features for each power grid device by using the standard data set based on a power grid device fault physical mechanism; the method comprises the following steps of: constructing a graph structure by taking power grid equipment as nodes and taking an electrical connection relationship or a geographic proximity relationship among the power grid equipment as edges, and each node in the graph structure comprises advanced features on each time slice; and inputting the advanced features into a trained space-time diagram neural network model, and outputting a dynamic fault probability value of each power grid device in a future time period by the space-time diagram neural network model. The method realizes the conversion from passive response to active early warning, and has the advantages of high prediction precision, strong real-time performance and wide applicability.

Description

Technical Field

[0001] This invention relates to a method and system for predicting the dynamic fault probability of power grid equipment under extreme weather conditions based on physical information and spatiotemporal deep learning, belonging to the field of power system digitalization and disaster prevention and mitigation technology. Background Technology

[0002] In modern society, the stability of the power system is crucial for ensuring social operation and economic development. With global climate change, extreme weather events such as typhoons, hailstorms, and extreme heat are becoming more frequent, posing a serious threat to the safe operation of the power grid. Accurately quantifying the probability of equipment failure under extreme weather conditions is key to the power grid's disaster prevention and mitigation, emergency dispatch, and risk early warning systems.

[0003] Existing power grid disaster prevention technologies for extreme weather suffer from the following limitations: Current methods heavily rely on historical fault data, estimating probabilities by statistically analyzing the number of faults in specific regions and under specific weather intensities. Their fatal flaw lies in their inability to cope with novel extreme weather events and their failure to consider dynamic factors such as equipment aging and current health status, resulting in lagging and crude predictions. Furthermore, in terms of model building, existing technologies often employ simplified mathematical models to describe complex fault mechanisms. For example, some studies attempt to use linear regression or logistic regression models to establish the relationship between meteorological parameters and equipment faults. While these methods can reflect the impact of single factors to some extent, they cannot effectively characterize the nonlinear characteristics of multi-factor coupling. In actual operation, equipment faults are often the result of multiple factors acting together, such as the superposition effect of wind and ice loads, and the interaction between lightning strikes and equipment insulation levels. These complex coupling relationships are difficult to accurately represent in existing models, and existing methods often focus only on one type of data, failing to effectively integrate and collaboratively analyze real-time equipment monitoring data, geographical environment data, and equipment ledger data, leading to insufficient information utilization and a single predictive dimension. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a method and system for predicting the dynamic failure probability of power grid equipment based on physical information and spatiotemporal deep learning under extreme weather conditions, so as to perform refined and dynamic failure probability prediction of power grid equipment under extreme weather events.

[0005] Prior to this invention, a method for predicting the dynamic fault probability of power grid equipment based on physical information and spatiotemporal deep learning under extreme weather conditions is provided, characterized by comprising:

[0006] Acquire multi-source heterogeneous data of power grid equipment within the target area, and perform unified spatiotemporal benchmark alignment and grid interpolation on the multi-source heterogeneous data to obtain a standard dataset;

[0007] Based on standard datasets, high-level features are constructed for each power grid device using the physical mechanisms of power grid equipment failures.

[0008] A graph structure is constructed using power grid equipment as nodes and electrical connections or geographical proximity relationships between power grid equipment as edges. Each node in the graph structure contains high-level features at each time slice. The high-level features include comprehensive operating condition coefficients, comprehensive mechanical stress functions, and comprehensive electrical stress functions.

[0009] By inputting high-level features into the trained spatiotemporal graph neural network model, predictions are made to obtain the future dynamic failure probability value of each power grid device.

[0010] Prior to this, the formula for calculating the comprehensive working condition factor is:

[0011] ,

[0012] in, The score is based on the service life of the power grid equipment; The score is based on the latest defect level. The score is based on the condition of the most recent maintenance. Scoring based on the service life of power grid equipment The weight, The weighting of the score based on the latest defect level, The weighting is based on the score of the most recent maintenance status. , and The sum is 1.

[0013] Preferably, the comprehensive mechanical stress function The calculation expression is:

[0014] ,

[0015] in, It is a function with maximum value. For wind load, For ice load, This is the preset coupling coefficient.

[0016] Preferably, the spatiotemporal graph neural network model includes an input layer, a time encoder, a spatial aggregator, and an output layer connected in sequence. After receiving high-level features of the power grid equipment, the time encoder captures the temporal evolution of the state of each equipment through a gated loop unit and outputs a feature hidden state containing temporal features. The spatial aggregator performs node feature transformation on the feature hidden state containing temporal features based on a graph attention network, calculates attention coefficients based on dynamic spatial influence factors, performs normalization processing, and performs weighted fusion of neighbor nodes to obtain neighbor information. The output layer maps the features that have fused spatiotemporal information into dynamic fault probability values ​​through a fully connected layer and a sigmoid function.

[0017] Firstly, a dynamic spatial influence factor is introduced into graph attention networks. Adjust attention coefficient;

[0018] in, The expression is:

[0019] ,

[0020] in, For nodes Place The meteorological characteristic vector includes wind speed and wind direction at any given moment; For nodes Place The meteorological characteristic vector includes wind speed and wind direction at any given moment; For nodes With nodes The angle between the line connecting the two points and the current wind direction; The learnable parameter matrix; This indicates a vector concatenation operation.

[0021] Prioritize training the spatiotemporal graph neural network model, including:

[0022] The spatiotemporal graph neural network model is trained using the overall loss function until it converges to a certain value.

[0023] The expression for the overall loss function is as follows:

[0024] ,

[0025] in For the prediction error loss term, These are preset weighted hyperparameters; This is the loss term for physical consistency constraints.

[0026] Preferably, the physical consistency constraint loss term includes at least a constraint term based on the comprehensive operating condition coefficient or a constraint term based on the comprehensive mechanical stress coefficient;

[0027] Among them, the constraint terms based on the comprehensive operating condition coefficient are:

[0028] ,

[0029] Constraints based on the comprehensive mechanical stress coefficient:

[0030] ,

[0031] in, This represents the total number of training samples in the standard dataset. For scaling functions, For power grid equipment The comprehensive mechanical stress function value, For power grid equipment The comprehensive working condition coefficient, The first preset threshold, The second preset threshold, Output power grid equipment for spatiotemporal graph neural network model The dynamic failure probability value.

[0032] Prioritize multi-source heterogeneous data, including geographic environment data, meteorological forecast data, and equipment ledger data.

[0033] Prioritized, including:

[0034] The data acquisition and integration module is used to acquire multi-source heterogeneous data from power grid equipment within the target area.

[0035] The data governance and storage module is used to perform unified spatiotemporal benchmark alignment and grid interpolation on multi-source heterogeneous data to obtain a standard dataset.

[0036] The analysis and computation engine module utilizes a standard dataset to construct high-level features for each power grid device based on the physical mechanisms of power grid equipment failures. It constructs a graph structure with power grid devices as nodes and electrical connections or geographical proximity relationships between devices as edges. Each node in the graph structure contains high-level features at each time slice. These high-level features include a comprehensive operating condition coefficient, a comprehensive mechanical stress function, and a comprehensive electrical stress function. The high-level features are then input into a trained spatiotemporal graph neural network model, which outputs the dynamic failure probability value for each power grid device in future time periods.

[0037] The results display module outputs the time-series dynamic fault probability of each power grid device in the form of a list or curve, and visualizes it on a geographic information map. Based on the preset color-coded warning standards, it identifies power grid devices and peak risk periods.

[0038] Preferably, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the methods described herein.

[0039] Preferably, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described herein.

[0040] The beneficial effects achieved by this invention are as follows:

[0041] First, this invention enables dynamic forward-looking early warning, transforming the disaster prevention and operation mode. Unlike traditional static estimation methods based on historical statistics, this invention utilizes a time-series deep learning model to process real-time and forecast data, generating a dynamic fault probability curve that continuously changes over multiple future time periods. This curve accurately reflects the evolution of risk with meteorological processes and precisely locates the peak risk period, thus providing a crucial decision-making window for power grid disaster prevention scheduling and emergency response, achieving a fundamental shift from passive accident response to proactive risk early warning.

[0042] Second, this invention deeply integrates physical mechanisms to enhance model interpretability and generalization ability. Addressing the common problem that data-driven models are often considered "black boxes" with insufficient generalization, this invention deeply integrates the physical mechanisms of failure into the model construction process. By constructing high-level features with clear physical meanings, such as "comprehensive operating condition coefficients" and "comprehensive mechanical stress functions," strong prior knowledge is injected into the model. These features are directly related to the causes of failure, not only enhancing the interpretability of the results but also effectively guiding the model to learn physical laws. This allows the model to maintain high prediction accuracy even in novel extreme weather scenarios not covered by training data, significantly improving practical reliability.

[0043] Third, this invention deeply integrates multi-source heterogeneous data, fully exploring the value of information synergy and enhancing the predictive dimension. This invention breaks down the barriers between meteorological, power grid, and geographic environmental data, achieving deep integration of multi-source heterogeneous data through unified spatiotemporal benchmark alignment and interpolation processing. The core advantage of the spatiotemporal graph neural network model lies in its ability to simultaneously mine the spatiotemporal correlation patterns inherent in these multi-source heterogeneous data: capturing the evolution trend of meteorological conditions and equipment status in the time dimension; and capturing the propagation and correlation characteristics of fault risks along electrical connections and geographic space through graph structure in the spatial dimension. This collaborative analysis of multi-dimensional information enables the model to quantify the complex coupling effects of multiple factors, which is impossible with traditional single-dimensional analysis methods.

[0044] In summary, this invention comprehensively improves the accuracy, foresight, refinement, and practicality of predicting the probability of power grid equipment failure under extreme weather conditions through multi-source information fusion, combination of physical mechanisms and deep learning, mining of spatiotemporal dual dependencies, and systematic engineering implementation, providing effective technical support for building a robust smart grid and improving power supply reliability. Attached Figure Description

[0045] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of the present invention.

[0047] Figure 2 This is a schematic diagram of the spatiotemporal graph neural network model structure of the present invention.

[0048] Figure 3 This is a flowchart illustrating the model training process in this invention. Detailed Implementation

[0049] See Figure 1 This embodiment uses the application of a power grid in a coastal area during typhoon weather as an example to demonstrate the specific implementation process of this method and system. Those skilled in the art will understand that the application during typhoon weather is merely an example, and the present invention is also applicable to other extreme weather scenarios such as hail, wildfires, and extreme heat.

[0050] S1: Data acquisition and preprocessing.

[0051] Collect geographical environment data, meteorological data, and ledger data of power grid equipment within the target area.

[0052] The meteorological data includes typhoon forecast data for the target area obtained from the meteorological bureau, including real-time monitoring data and forecast data for the next 0-72 hours, such as wind speed, wind direction, temperature, humidity, air pressure, precipitation, and lightning location information.

[0053] The ledger data includes ledger data for equipment such as transmission lines and towers obtained from the production management system, including equipment ID, latitude and longitude coordinates, model, commissioning date, and design wind resistance level; and the latest health status score, known defects, and maintenance history of power grid equipment obtained from the defect database and inspection records.

[0054] The geographic environment data includes altitude, slope, and aspect data of the location of the power grid equipment obtained from the geographic information system, vegetation type and average height around the power line corridor obtained from the forestry department's data interface, and soil moisture data obtained from remote sensing satellite inversion data.

[0055] Considering that the impact of extreme weather on power grid equipment is the result of the combined effects of meteorological conditions, geographical environment, and equipment inventory data, introducing geographical environment data and inventory data as supplements to fault probability prediction is beneficial to improving the accuracy of prediction results.

[0056] After acquiring multi-source heterogeneous data, including geographic environment data, meteorological data, and ledger data, preprocessing was performed:

[0057] Using Beijing time as the time base, the timestamps of all geographic environmental data, meteorological data, and ledger data are unified. Using the WGS-84 coordinate system as the spatial base, all power grid equipment, meteorological grid points, and geographic information are unified under this coordinate system.

[0058] Since the acquired meteorological data is gridded data, a bilinear interpolation method is used to interpolate the wind speed, wind direction and other data of the four vertices of the grid where each power grid equipment coordinate point is located to the specific coordinate point of the equipment, thereby generating time-series meteorological data that corresponds one-to-one with each equipment.

[0059] The missing and outlier values ​​in the geographic environment data, time-series meteorological data, and ledger data are checked and processed, and the numerical features are standardized to form a standard dataset.

[0060] S2: Advanced feature construction.

[0061] In this embodiment, a high-level feature is constructed for each power grid device at each future forecast time t based on a standard dataset and physical mechanisms. Taking tower equipment as an example, the following characteristics are mainly calculated:

[0062] Comprehensive operating condition coefficient: Used to quantify the vulnerability of power grid equipment itself. The higher the value, the better the equipment's health condition and the stronger its risk resistance. In this embodiment, the expression for the comprehensive operating condition coefficient is:

[0063] ,

[0064] in The scoring is based on the service life of power grid equipment, assuming an annual aging and depreciation rate of 3%; The scoring is based on the latest defect level, with no defect = 1.0, minor defect = 0.8, major defect = 0.5, and critical defect = 0.2. The score is based on the most recent maintenance status; 1.2 is given if maintenance was carried out within the last 3 months, otherwise 1.0; Scoring based on the service life of power grid equipment weights, For the weighting of scores based on the latest defect level, The weighting is based on the score of the most recent maintenance status. , and The sum is 1.

[0065] Comprehensive mechanical stress function: quantifying the combined effect of wind load and potential ice load, in this embodiment, its expression is: ,in It is a function with maximum value. For wind load, For ice load, coefficient Very small This indicates that under extreme conditions such as rain, snow, and ice, the effects of wind and ice can amplify each other, drastically amplifying mechanical stress and accurately capturing the most dangerous working conditions. This embodiment uses a typhoon as an example and does not consider icing, therefore ice load... = 0, coefficient The value is 0.1.

[0066] Comprehensive electrical stress function: This function quantifies the destructive intensity of wildfires, lightning strikes, etc., on the external insulation of equipment. In this embodiment, the expression is:

[0067]

[0068] in For wildfire risk function, This is the lightning impact function. In this embodiment, wildfires are not considered; the wildfire risk function... .

[0069] Tree and bamboo lodging risk index: based on wind speed Vegetation height and soil moisture The calculation, in this embodiment, is expressed as:

[0070] ,

[0071] The higher the index, the greater the risk that trees or bamboo near the line corridor will fall and damage the line under strong winds and moist soil conditions.

[0072] The features calculated above are combined with other standardized original features to form each device. At any moment Advanced features .

[0073] S3: Graph Structure Construction

[0074] In this embodiment, the power poles are used as nodes according to the power grid topology. The lines connecting the towers are used as edges to construct a graph. If two electrical grid devices are directly connected, then an edge is established between them. .

[0075] S4: Dynamic probability calculation.

[0076] In this embodiment, a pre-trained spatiotemporal graph neural network model is used to perform calculations on the multi-source heterogeneous data after preprocessing and high-level feature construction.

[0077] High-level timing characteristics of each node device Input gated loop unit. Update the feature hiding state according to its internal mechanism, and finally output the feature representation that incorporates the historical information of the power grid equipment itself. In this embodiment, its expression is:

[0078] ,

[0079] in This is the forward computation process of the gated loop unit.

[0080] Spatial aggregation operations are performed to capture spatial dependencies. A graph attention network is used as the spatial aggregator, and an attention mechanism is introduced to calculate the association weights between a node and its neighboring nodes. The specific implementation process includes:

[0081] The first step is to represent the temporal features of each node. Perform a shared linear transformation by multiplying by a learnable weight matrix. To improve the expressive power of the model and obtain the transformed feature representation. .

[0082] The second step is to calculate the nodes. Each of its connected nodes Unnormalized attention coefficients between A single-layer feedforward neural network is used and through The activation function introduces nonlinearity, and its calculation formula is as follows: ,in It is a learnable attention vector. This indicates a vector concatenation operation.

[0083] The third step involves introducing a dynamic spatial influence factor driven by real-time meteorological conditions into the graph attention network. This is used to correct the attention weights between nodes, enabling the spatiotemporal graph neural network model to adapt to the rapidly changing spatial dependencies under extreme weather conditions. The expression is as follows:

[0084] ,

[0085] in, For nodes Place The meteorological characteristic vector includes wind speed and wind direction at any given moment; For nodes Place The meteorological characteristic vector includes wind speed and wind direction at any given moment; For nodes With nodes The angle between the line connecting the two points and the current wind direction; The learnable parameter matrix; This indicates a vector concatenation operation.

[0086] The fourth step is to use the softmax function to normalize the attention coefficients of all neighbors, thus obtaining the final normalized attention weights. ,

[0087] in, It is a node The set of all neighboring nodes, It is a natural exponential function; the weighting coefficient The higher the value, the stronger the neighboring node. The state of the central node The greater the impact of the failure risk.

[0088] The fifth step involves using the normalized attention weights as coefficients to perform a weighted summation of the transformed features of neighboring nodes, and then applying this summation through a non-linear activation function. The transformation is performed to generate a new node feature representation that incorporates spatial information. In practical applications, the above spatial aggregation process can be iterated multiple times, enabling each node to receive and merge information from multi-hop neighbors, thereby capturing deeper and more complex spatial dependency patterns in the graph.

[0089] Feature representation Nodes are obtained through a fully connected layer. The final feature representation after processing by the L-layer spatiotemporal network Then, the Sigmoid activation function is used to map the high-level feature representation after spatiotemporal encoding and aggregation to a fault probability value between 0 and 1. .

[0090] ,in and These are the weights and biases of the output layer. This is the Sigmoid function. The model outputs the hourly dynamic failure probability values ​​for all devices over the next 72 hours.

[0091] S5: Dynamic risk visualization and output.

[0092] The application demonstration and early warning module receives the fault probability calculation results sent by the analysis and calculation engine module.

[0093] In terms of visualization, the system uses a GIS map platform for dynamic rendering: a heat map clearly displays the distribution of fault risks across all devices, and a four-level color-coded warning standard is defined based on preset thresholds—green indicates a safe state, yellow indicates a low-risk warning, orange indicates a medium-risk warning, and red indicates a high-risk alarm. Operators can click on any device node on the map to further view the fault probability change curve of that device over a specific future period, thereby accurately grasping the trend of risk evolution.

[0094] Furthermore, in this embodiment, the training process of the spatiotemporal graph neural network model introduces a physical consistency constraint loss term on top of the traditional supervised learning loss. This term penalizes prediction results that violate the physical mechanism of equipment failure, thereby improving the model's generalization ability and physical interpretability. For details, please refer to [link / reference]. Figure 3 The spatiotemporal graph neural network model minimizes the overall loss function. The overall loss function expression obtained through training is as follows: .

[0095] in The prediction error loss term is used to represent the difference between the model's predicted probability output by the spatiotemporal graph neural network model and the actual fault label. These are preset weighted hyperparameters; The physical consistency constraint loss term is used to compare the model output with the expected result derived from the physical mechanism of the failure and to penalize the predicted output that violates physical common sense.

[0096] In this embodiment, the physical consistency constraint loss term It includes constraint terms based on the comprehensive working condition coefficient and constraint terms based on the comprehensive mechanical stress coefficient, and its expression is: ,

[0097] in, The total number of training samples, For a scaling function, For power grid equipment The comprehensive mechanical stress function value, For power grid equipment The comprehensive working condition coefficient, The first preset threshold, The second preset threshold, Output power grid equipment for spatiotemporal graph neural network model The probability of failure.

[0098] In this embodiment of the application, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the methods described above.

[0099] In this application embodiment, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0100] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0101] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention described herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not invented herein. The specification and embodiments are to be considered exemplary only.

[0102] The above specific embodiments further illustrate the purpose, technical solution and beneficial effects of this application. It should be understood that the above are only specific embodiments of this application and are not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of this application should be included within the scope of protection of this application.

Claims

1. A method for predicting the dynamic fault probability of power grid equipment under extreme weather conditions based on physical information and spatiotemporal deep learning, characterized in that, include: Acquire multi-source heterogeneous data of power grid equipment within the target area, and perform unified spatiotemporal benchmark alignment and grid interpolation on the multi-source heterogeneous data to obtain a standard dataset; Based on standard datasets, high-level features are constructed for each power grid device using the physical mechanisms of power grid equipment failures. A graph structure is constructed using power grid equipment as nodes and electrical connections or geographical proximity relationships between power grid equipment as edges. Each node in the graph structure contains high-level features at each time slice. The high-level features include comprehensive operating condition coefficients, comprehensive mechanical stress functions, and comprehensive electrical stress functions. By inputting high-level features into the trained spatiotemporal graph neural network model, predictions are made to obtain the future dynamic failure probability value of each power grid device.

2. The method for predicting the dynamic fault probability of power grid equipment under extreme weather conditions based on physical information and spatiotemporal deep learning, as described in claim 1, is characterized in that... The formula for calculating the comprehensive working condition coefficient is as follows: ， in, The score is based on the service life of the power grid equipment; The score is based on the latest defect level. The score is based on the condition of the most recent maintenance. Scoring based on the service life of power grid equipment The weight, The weighting of the score based on the latest defect level, The weighting is based on the score of the most recent maintenance status. , and The sum is 1.

3. The method for predicting the dynamic fault probability of power grid equipment under extreme weather conditions based on physical information and spatiotemporal deep learning, as described in claim 1, is characterized in that... The comprehensive mechanical stress function The calculation expression is: ， in, It is a function with maximum value. For wind load, For ice load, This is the preset coupling coefficient.

4. The method for predicting the dynamic fault probability of power grid equipment under extreme weather conditions based on physical information and spatiotemporal deep learning, as described in claim 1, is characterized in that... The spatiotemporal graph neural network model comprises an input layer, a time encoder, a spatial aggregator, and an output layer connected in sequence. After receiving high-level features of the power grid equipment, the time encoder captures the temporal evolution of the state of each device through a gated recurrent unit and outputs a feature hidden state containing temporal features. The spatial aggregator performs node feature transformation on the feature hidden state containing temporal features based on a graph attention network, calculates attention coefficients based on dynamic spatial influence factors, performs normalization processing, and weighted fusion of neighbor nodes to obtain neighbor information. The output layer maps the features that have fused spatiotemporal information into dynamic fault probability values ​​through a fully connected layer and a sigmoid function.

5. The method for predicting the dynamic fault probability of power grid equipment under extreme weather conditions based on physical information and spatiotemporal deep learning, as described in claim 1, is characterized in that... Introducing dynamic spatial influence factors into graph attention networks Adjust attention coefficient; in, The expression is: ， in, For nodes Place The meteorological characteristic vector includes wind speed and wind direction at any given moment; For nodes Place The meteorological characteristic vector includes wind speed and wind direction at any given moment; For nodes With nodes The angle between the line connecting the two points and the current wind direction; The learnable parameter matrix; This indicates a vector concatenation operation.

6. The method for predicting the dynamic fault probability of power grid equipment under extreme weather conditions based on physical information and spatiotemporal deep learning, as described in claim 1, is characterized in that... The spatiotemporal graph neural network model has been trained and includes: The spatiotemporal graph neural network model is trained using the overall loss function until it converges to a certain value. The expression for the overall loss function is as follows: ， in For the prediction error loss term, These are preset weighted hyperparameters; This is the loss term for physical consistency constraints.

7. A method for predicting the dynamic fault probability of power grid equipment under extreme weather conditions based on physical information and spatiotemporal deep learning, as described in claim 6, is characterized in that... The physical consistency constraint loss term includes at least a constraint term based on the comprehensive operating condition coefficient or a constraint term based on the comprehensive mechanical stress coefficient; Among them, the constraint terms based on the comprehensive operating condition coefficient are: ， Constraints based on the comprehensive mechanical stress coefficient: ， in, This represents the total number of training samples in the standard dataset. For scaling functions, For power grid equipment The comprehensive mechanical stress function value, For power grid equipment The comprehensive working condition coefficient, The first preset threshold, The second preset threshold, Output power grid equipment for spatiotemporal graph neural network model The dynamic failure probability value.

8. The method for predicting the dynamic fault probability of power grid equipment under extreme weather conditions based on physical information and spatiotemporal deep learning, as described in claim 1, is characterized in that... Multi-source heterogeneous data includes geographic environment data, meteorological forecast data, and equipment ledger data.

9. A dynamic fault probability prediction system for power grid equipment based on physical information and spatiotemporal deep learning under extreme weather conditions, as described in claim 1, is characterized in that... include: The data acquisition and integration module is used to acquire multi-source heterogeneous data from power grid equipment within the target area. The data governance and storage module is used to perform unified spatiotemporal benchmark alignment and grid interpolation on multi-source heterogeneous data to obtain a standard dataset. The analysis and computation engine module is used to construct high-level features for each power grid device based on the physical mechanism of power grid equipment failure using standard datasets. A graph structure is constructed using power grid equipment as nodes and electrical connections or geographical proximity relationships between power grid equipment as edges. Each node in the graph structure contains high-level features at each time slice. The high-level features include comprehensive operating condition coefficients, comprehensive mechanical stress functions, and comprehensive electrical stress functions. The high-level features are input into a trained spatiotemporal graph neural network model, which outputs the dynamic failure probability value of each power grid equipment in future time periods. The results display module outputs the time-series dynamic fault probability of each power grid device in the form of a list or curve, and visualizes it on a geographic information map. Based on the preset color-coded warning standards, it identifies power grid devices and peak risk periods.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 7.

Images