A Typhoon Disaster Risk Early Warning Method Based on Fuded Data and Meteorological Big Model
By integrating data with a large meteorological model, a sequence of equipment risk characteristics is constructed and a graph neural network is used to solve the problems of continuous risk description and dynamic evolution at the equipment level in existing typhoon disaster risk early warning methods, thereby achieving high-precision risk prediction and proactive prevention and control.
Patent Information
- Application Number
- CN202611610499.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-19
- Publication Date
- 2026-11-17
AI Technical Summary
Existing typhoon disaster risk early warning methods are unable to comprehensively consider meteorological, geographical, power grid topology, equipment status and historical disaster information, cannot achieve continuous risk description at the equipment level, have limited risk prediction accuracy and generalization ability, are difficult to dynamically evolve and predict, and lack sufficient connection between early warning and proactive prevention and control.
By acquiring typhoon meteorological, power grid equipment, spatial environment, and historical fault data, and utilizing meteorological big data models and graph neural networks, a sequence of equipment risk characteristics is constructed. Combined with the power grid topology map, risk prediction is performed, and typhoon disaster risk warning information is generated, including equipment risk level and warning time.
It enables continuous risk description at the equipment level, improves the accuracy of risk prediction and dynamic evolution capability, enhances the connection between early warning and proactive prevention and control, and allows for advance planning of prevention and control measures.
Smart Images

Figure CN123117841A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of wind disaster early warning technology, specifically involving a typhoon disaster risk early warning method based on fused data and a meteorological big data model. Background Technology
[0002] Typhoons are a significant natural disaster affecting the safe operation of the Shanghai power grid, characterized by their wide impact range, long duration, high destructive intensity, and high uncertainty in evolution. During super typhoons, the coupled effects of multiple disaster factors, such as strong winds, torrential rain, thunderstorms, and urban flooding, can easily lead to tower collapse, line tripping, equipment malfunctions, and large-scale power outages. With the expansion of urban scale and the increasing complexity of power grid structures, traditional risk assessment methods relying on empirical rules are insufficient to meet the needs of refined disaster prevention under extreme weather conditions. There is an urgent need to integrate multi-source heterogeneous data to achieve dynamic risk perception and accurate assessment. In recent years, the development of artificial intelligence, big data, and meteorological big data modeling technologies has provided new approaches to building intelligent prediction models for power grid disasters. By integrating data from meteorology, power grid operation, equipment failure, and the spatial environment, it is possible to identify the typhoon impact range, predict equipment risks, and analyze disaster trends, providing support for proactive power grid defense and emergency decision-making.
[0003] Existing typhoon disaster risk early warning methods mainly include meteorological rule assessment, statistical model prediction, and machine learning data-driven methods. Traditional methods typically construct risk discrimination rules based on meteorological indicators such as typhoon level, wind speed, and rainfall. While simple to implement, these methods struggle to account for complex environmental factors and equipment differences. Subsequently, methods such as logistic regression, support vector machines (SVM), random forests, and gradient boosting trees (GBDT) have been used to uncover the correlation between meteorological factors and equipment failures, enabling disaster risk prediction. With the development of deep learning technology, models such as convolutional neural networks (CNN), recurrent neural networks (RNN), and long short-term memory networks (LSTM) are increasingly being applied to the spatiotemporal correlation analysis of disasters. Simultaneously, large-scale meteorological artificial intelligence models, by learning from extensive historical meteorological data, have improved the ability to identify and refine extreme weather events, providing new technical support for power grid disaster risk early warning.
[0004] For example, Chinese patent application CN115688032A discloses a method and system for power grid risk early warning under typhoon disasters based on multi-source data fusion. This method divides and matches power grid data, geographic data, and meteorological data into grids according to time and spatial dimensions, establishes vulnerability curves for power grid components within the grids, uses Copeland scores to characterize the overall vulnerability of the power grid within the corresponding grids, and trains a regression decision tree based on the gridded risk early warning dataset to predict the number of power grid failures according to typhoon forecast data.
[0005] However, this method organizes multi-source data using time units and spatial grids, and primarily predicts the number of historical faults of power grid components within the grid. Grid aggregation tends to weaken the individual differences and state changes of different devices, making it difficult to form a continuous risk description at the device level. The method characterizes grid meteorological conditions using wind speed at the grid's geometric center, which is insufficient to reflect local meteorological differences caused by typhoon path changes, wind and rain coupling, and complex spatial environments. Its predictions, based on vulnerability curves, Copeland scores, and regression decision trees, fail to fully characterize the time-varying nonlinear relationships between meteorological factors, device status, and spatial environmental factors. Furthermore, the geographical proximity of the spatial grid does not perfectly align with the actual electrical connections of the devices, making it difficult to reflect the associated impact of fault risks along electrical connection paths. This method mainly outputs the predicted number of power grid faults, making it difficult to identify specific high-risk devices and their continuous risk changes and warning times, thus offering limited support for refined device-level early warning.
[0006] Existing technologies for power grid risk early warning applications during super typhoons still have the following shortcomings: First, they lack the ability to fuse multi-source heterogeneous data. Existing methods mostly rely on single meteorological variables or limited operational data, making it difficult to comprehensively consider meteorological, geographical, power grid topology, equipment status, and historical disaster information to fully characterize the disaster impact mechanism. Second, their risk prediction accuracy and generalization ability are limited. Traditional models rely on manual feature construction and are insufficient in modeling the complex nonlinear relationships in typhoon path changes, wind and rain coupling, and disaster propagation. Third, their ability to predict the dynamic evolution of disasters is insufficient. Most methods focus on single-moment risk assessment or regional fault quantity prediction, making it difficult to achieve continuous dynamic prediction of the risk of specific power grid equipment as the typhoon evolves. Fourth, the connection between early warning and proactive prevention and control is insufficient. Existing systems mostly remain at the level of risk level identification or fault quantity statistics, lacking intelligent decision support capabilities for inspection, emergency repair, material allocation, and operational adjustments. Summary of the Invention
[0007] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a typhoon disaster risk early warning method based on fused data and meteorological big data model.
[0008] The objective of this invention can be achieved through the following technical solutions: This invention provides a typhoon disaster risk early warning method based on fused data and a meteorological big data model, comprising the following steps: Acquire typhoon meteorological data, power grid equipment data, spatial environment data, and historical fault data of the target power grid; and correlate the power grid equipment data, spatial environment data, and historical fault data based on equipment identification and time to obtain equipment correlation data. The typhoon meteorological data is input into the meteorological big model to obtain the typhoon meteorological spatiotemporal characteristics. The typhoon meteorological spatiotemporal characteristics are then fused with the equipment-related data based on the equipment location and time to obtain the equipment risk characteristic sequence of each power grid equipment. A power grid topology map is constructed based on the power grid equipment and the electrical connections between the power grid equipment. The risk feature sequence of the equipment is input into a graph neural network risk prediction model based on the power grid topology map to obtain the future risk sequence of each power grid equipment. Based on the future risk sequence, the risk level and warning time of each power grid device are determined, and typhoon disaster risk warning information is generated.
[0009] Furthermore, the acquisition of typhoon meteorological data, power grid equipment data, spatial environment data, and historical fault data of the target power grid specifically includes: Gridded meteorological forecast data and measured meteorological data, including precipitation, temperature, humidity, typhoon center location, typhoon intensity, typhoon movement path and wind speed, are obtained as the typhoon meteorological data; Acquire basic ledgers and operating status data of power transmission and distribution equipment, including equipment location, line length, tower type, equipment commissioning years, equipment health status, load level, and online monitoring data, as the power grid equipment data; Obtain geospatial data including geographic information system data, digital elevation model data, land use types, river and road distribution, and urban building distribution, as the spatial environment data; Acquire equipment failure records, emergency repair work orders, meteorological disaster process data, and disaster loss data corresponding to each typhoon disaster, and form historical failure samples according to the correspondence between typhoon characteristics, equipment status, and failure results, as the historical failure data.
[0010] Furthermore, the process of associating the power grid equipment data, spatial environment data, and historical fault data based on equipment identification and time to obtain equipment association data specifically includes: The integrity of the typhoon meteorological data, power grid equipment data, spatial environment data and historical fault data is checked. Time interpolation is performed on the missing meteorological time series, and inverse distance weighted spatial compensation is performed on the missing gridded meteorological data. The missing equipment operating status is repaired by combining the electrical connection relationship between power grid equipment and historical operating data. The missing data is input into the isolated forest model to obtain the anomaly score of each data sample. Data samples with anomaly scores exceeding the preset anomaly threshold are identified as anomalous data. The anomalous data is then verified based on the operating status of the adjacent equipment of the power grid equipment corresponding to the anomalous data. The data that has completed the abnormal data verification is subjected to coordinate transformation, timestamp synchronization and data format standardization. The spatial environment data is spatially matched with the corresponding power grid equipment data according to the equipment location. The historical fault data is associated with the corresponding power grid equipment data according to the equipment identifier and timestamp. Data on power grid equipment, spatial environment, and historical faults that have the same device identifier and are located within the same time window are combined into a single device association record. These device association records are then arranged in chronological order to obtain the device association data.
[0011] Furthermore, the fusion of the typhoon meteorological spatiotemporal characteristics with the associated equipment data based on equipment location and time yields a sequence of equipment risk characteristics for each power grid device, specifically including: From the aforementioned typhoon meteorological spatiotemporal characteristics, wind field characteristics, rainfall and urban flooding characteristics, and typhoon spatial impact characteristics are extracted. The wind field characteristics include maximum sustained wind speed, gust speed, wind speed change rate, and the angle between wind direction and line direction. The rainfall and urban flooding characteristics include cumulative rainfall, maximum rainfall intensity, continuous rainfall duration, and soil moisture index. The typhoon spatial impact characteristics include the distance between the typhoon center and the equipment location, the wind circle coverage status, the typhoon's movement direction, and the future impact time. The equipment vulnerability features and spatial exposure features are extracted from the equipment association data. The equipment vulnerability features include the equipment's years of operation, number of historical failures, attributes of important power supply nodes, load level, and maintenance status. The spatial exposure features include terrain elevation, slope, distance from the river, urban building density, and underground space distribution. Based on the historical fault data, a disaster knowledge graph containing the relationships between typhoon type, meteorological conditions, equipment type and fault mode is constructed, and historical disaster association features are extracted from the disaster knowledge graph based on the entities and relationships corresponding to each power grid equipment. Based on the location and time of the equipment, the wind field characteristics, rainfall and waterlogging characteristics, typhoon spatial impact characteristics, equipment vulnerability characteristics, spatial exposure characteristics, and historical disaster association characteristics are matched with the corresponding power grid equipment. The characteristics of the same power grid equipment at multiple consecutive times are arranged in chronological order to obtain the equipment risk characteristic sequence.
[0012] Furthermore, the step of inputting the typhoon meteorological data into the meteorological big data model to obtain the spatiotemporal characteristics of typhoon meteorology specifically includes: Typhoon meteorological data from multiple consecutive moments are combined into a meteorological input sequence in chronological order. The meteorological variables, temporal location, and spatial grid location in the meteorological input sequence are encoded into meteorological feature vectors through an embedding layer. The meteorological feature vector is input into a large meteorological model based on the Transformer architecture. The correlation features between different meteorological variables, different time locations, and different spatial grid locations are extracted through a multi-head self-attention layer. The correlation features are then iteratively encoded through multiple Transformer encoders connected in sequence. The output layer of the meteorological model outputs the typhoon path characteristics, wind field diffusion characteristics, and rainfall evolution characteristics corresponding to multiple future time points, and combines the typhoon path characteristics, wind field diffusion characteristics, and rainfall evolution characteristics according to time and spatial grid positions to obtain the typhoon meteorological spatiotemporal characteristics. The meteorological model is pre-trained using historical meteorological data and adapted using historical typhoon meteorological data of the target power grid area.
[0013] Furthermore, the fusion of the typhoon meteorological spatiotemporal characteristics with the associated equipment data based on equipment location and time yields a sequence of equipment risk characteristics for each power grid device, specifically including: Based on the equipment location and time, extract the relevant information about the power grid equipment from the spatiotemporal characteristics of the typhoon meteorological data. At any moment Corresponding meteorological characteristics The device features are obtained by feature encoding the power grid equipment data, spatial environment data, and historical fault data in the device-related data. Spatial environment characteristics and historical fault characteristics ; The device features Spatial environment characteristics and historical fault characteristics Concatenate and encode into power grid features And determine the gating coefficient: Regarding the meteorological characteristics and the characteristics of the power grid field Perform gating fusion to obtain equipment At any moment Corresponding equipment risk characteristics: in, Indicates device At any moment The corresponding gating coefficient; This represents the Sigmoid activation function; Represents the gate weight matrix; Represents the gate bias vector; square brackets indicate feature concatenation; Indicates meteorological characteristics; Indicates equipment characteristics; Indicates spatial environmental characteristics; Indicates historical fault characteristics; Indicates characteristics of the power grid sector; Indicates the risk characteristics of the equipment; This indicates element-wise multiplication; The equipment risk characteristics of the same power grid equipment at multiple consecutive times Arranged in chronological order, the sequence of risk characteristics of the equipment is obtained.
[0014] Furthermore, the step of constructing a power grid topology map based on the power grid equipment and the electrical connections between the equipment, and inputting the equipment risk feature sequence into a graph neural network risk prediction model based on the power grid topology map to obtain the future risk sequence of each power grid equipment, specifically includes: The power grid equipment is set as a node in the power grid topology graph, the electrical connection relationship between different power grid equipment is set as an edge in the power grid topology graph, and the node characteristics of the corresponding node are set according to the equipment risk characteristic sequence corresponding to each power grid equipment. The device risk feature sequences of each node are encoded in a time dimension to obtain the initial node state of each node at different times. The node states are then transmitted and aggregated between adjacent nodes using the graph neural network risk prediction model. Wherein, the first... The node state of a layer is defined by the following formula: in, Represents device node At any moment The corresponding number Layer node status; Represents device node At any moment The corresponding number Layer node status; Indicates adjacent device nodes At any moment The corresponding number Layer node status; Indicates the relationship with device nodes A set of adjacent device nodes that are electrically connected; This represents the weight matrix corresponding to the node's own state. The weight matrix represents the state of adjacent nodes; Represents a non-linear activation function; The status of each device node that has completed the node status update is decoded according to future time points to obtain the future risk sequence of each power grid device at multiple consecutive future time points.
[0015] Furthermore, the graph neural network risk prediction model is trained under supervised supervision using the historical fault data, specifically including: The equipment risk feature sequence and power grid topology map corresponding to each power grid device during historical typhoons are used as training inputs. The corresponding power grid devices are used as training labels for whether they have failed, the time of failure, and the location of failure propagation during the historical typhoons. Risk classification loss, time prediction loss, and spatial prediction loss are used to jointly optimize the graph neural network risk prediction model. The device that outputs the trained graph neural network risk prediction model In the future The probability of an internal failure is defined as the equipment failure probability. The equipment is determined based on the power supply user level, load size, attributes of important power nodes, and the scope of fault impact. Equipment importance ; Determine equipment In the future Corresponding equipment risk index: in, Indicates device In the future The corresponding equipment risk index; Indicates device In the future The probability of equipment failure occurring within the system; Indicates device The importance of the equipment; The future risk sequence is formed by arranging the equipment risk indices corresponding to the same power grid equipment at multiple consecutive future moments in chronological order.
[0016] Furthermore, determining the risk level and early warning time of each power grid device based on the future risk sequence specifically includes: Acquire equipment Equipment location And the future Corresponding typhoon center location Determine the equipment Spatial distance from the center of the typhoon: Based on the spatial distance and future time The corresponding typhoon wind circle radius determines the equipment. In the future The corresponding direct impact of the typhoon: in, Indicates device The location of the equipment; Representing the equipment The horizontal and vertical coordinates; Indicates future time The corresponding location of the typhoon's center; These represent the horizontal and vertical coordinates of the typhoon's center location, respectively. Indicates device Spatial distance from the center of the typhoon; Indicates future time The corresponding radius of the typhoon's wind circle; Indicates device In the future The corresponding degree of direct impact of the typhoon; Based on the degree of direct impact of the typhoon, the angle between the wind direction and the route, and the terrain features corresponding to the equipment location, the equipment risk index in the future risk sequence is corrected to obtain the corrected future risk sequence. The equipment risk index in the corrected future risk sequence is compared with the preset risk level threshold to determine the risk level of each power grid equipment at different future times. The future time when the equipment risk index first reaches the warning threshold is determined as the warning time of the corresponding power grid equipment. The time difference between the warning time and the current time is determined as the warning lead time.
[0017] Furthermore, the generation of typhoon disaster risk warning information specifically includes: Based on the future risk sequence, risk level and warning time of each power grid equipment, generate risk area distribution data, high-risk equipment list, expected failure time data, equipment risk ranking data and future risk trend data; The risk level, warning time, power grid equipment data, and risk area distribution data are input into the prevention and control strategy recommendation model to construct a state vector containing the current typhoon status, power grid operation status, and risk distribution status. Equipment inspection, windproof reinforcement, operation mode adjustment, and emergency material deployment are set as candidate prevention and control actions. The action evaluation value is determined based on the expected disaster loss, execution cost, and power grid safety status corresponding to the candidate prevention and control actions. Based on the action evaluation value, the target prevention and control action corresponding to each high-risk equipment is determined from the candidate prevention and control actions, and the target prevention and control action is written into the typhoon disaster risk early warning information; The typhoon disaster risk warning information is sent to the power grid dispatching platform, emergency command platform and mobile terminal. The actual equipment failure data, actual failure time and actual prevention and control results after the typhoon disaster are written into the historical failure data. The model parameters of the graph neural network risk prediction model and the prevention and control strategy data of the prevention and control strategy recommendation model are updated.
[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) Existing methods often use the number of faults in a spatial grid or region as the prediction object, which easily weakens the differences in the operating years, health status, load level and historical faults of different equipment in the same region. This invention correlates and integrates meteorological, power grid equipment, spatial environment and historical fault data based on equipment identification, equipment location and time, and refines the risk analysis to specific power grid equipment and continuous time, thereby reducing the risk averaging caused by grid aggregation and improving the precision of high-risk equipment location.
[0019] (2) Existing methods typically use a limited number of meteorological indicators such as wind speed, rainfall, and typhoon level for risk assessment. These methods struggle to describe the temporal and spatial coupling evolution of factors such as typhoon path deflection, intensity changes, wind field diffusion, and continuous rainfall. This can easily lead to an oversimplification of the disaster effects in different regions and at different times during the same typhoon. This invention inputs typhoon meteorological data from multiple consecutive moments into a large meteorological model based on the Transformer architecture. Through a multi-head self-attention mechanism, it extracts the correlation features between different meteorological variables, different time locations, and different spatial grids to obtain typhoon path characteristics, wind field diffusion characteristics, and rainfall evolution characteristics for multiple future moments. It then uses historical typhoon data from the target power grid area for migration and adaptation. This transforms discrete meteorological indicators into meteorological spatiotemporal features with temporal continuity and spatial correlation, enabling equipment risk prediction to be dynamically adjusted as the typhoon moves, intensifies, weakens, and changes in wind and rain areas. This reduces the misjudgment of local risks caused by using only single-point wind speed or meteorological conditions at the grid center.
[0020] (3) Existing multi-source data fusion typically employs simple feature splicing or fixed weights, making it difficult to dynamically distinguish the actual contribution of meteorological factors, equipment vulnerability, and spatial exposure conditions to risk based on the stage of the equipment and the typhoon. For example, when a typhoon is not yet approaching, equipment aging and health status may play a major role, while when strong winds and heavy rains rapidly intensify, the impact of meteorological factors on short-term risk changes will significantly increase. This invention extracts features of wind field, rainfall and waterlogging, typhoon spatial impact, equipment vulnerability, and spatial exposure, and uses historical fault data to construct a disaster knowledge graph to obtain historical disaster association features. Furthermore, it dynamically integrates meteorological features and power grid domain features through gating. This allows for adaptive adjustment of the contribution of features from different sources to risk representation based on different equipment and different times, avoiding the obscuring of individual differences in equipment by a single meteorological condition, and also avoiding the weakening of the model's response to sudden meteorological changes by static equipment attributes, so that different equipment can still form differentiated risk results under the same meteorological environment.
[0021] (4) Existing technologies typically use geographical proximity to describe the risk associations between devices. However, geographical proximity does not necessarily imply actual electrical connections between devices, potentially leading to erroneous risk propagation between unrelated devices and overlooking the impact of associated faults along the actual power supply path. This invention constructs a power grid topology map based on the actual electrical connections between power grid devices, using devices as nodes and electrical connections as edges. The device risk feature sequences are input into a graph neural network as node features. The future risk sequence of each device is obtained through state transfer and aggregation between adjacent nodes. Therefore, the future risk of a device not only reflects its own typhoon impact and operating status but also integrates the states of adjacent devices with actual electrical connections, enabling risk associations to be expressed along the actual electrical path. Compared to establishing device relationships solely based on spatial distance, this reduces false associations between geographically adjacent but electrically independent devices and enhances the ability to identify fault propagation along the power supply path and continuous high-risk device groups.
[0022] (5) Existing models usually only predict whether the equipment will fail or the number of failures in a region, and it is difficult to take into account the time of failure and the location of its propagation. This invention uses the equipment failure status, the time of failure and the location of its propagation for joint training, and combines the equipment failure probability and the equipment importance to form an equipment risk index, so that the risk results simultaneously reflect the possibility of equipment failure and the importance of the consequences of failure, and avoid underestimating key power supply equipment by only ranking by failure probability.
[0023] (6) In view of the problem that pure data-driven models are prone to risk bias in scenarios with few samples such as extreme typhoon paths, special wind directions and complex terrain, this invention further corrects the future risk sequence based on the distance between the equipment and the typhoon center, the radius of the typhoon wind circle, the angle between the wind direction and the route, and the terrain features. In this way, the results of the model are constrained by the typhoon action conditions with actual physical meaning, thereby improving the rationality of equipment risk judgment in local complex environments.
[0024] (7) Existing methods mostly focus on predicting risk levels or the number of regional faults at a single moment. Even if it is possible to determine that equipment is at risk, it is difficult to determine when the risk will reach a level that requires action. This invention outputs a risk sequence for multiple consecutive future moments for each power grid device, and determines the future moment when the equipment risk index first reaches the warning threshold as the warning moment, thereby obtaining a lead time for warnings. This allows inspections, reinforcement, operation mode adjustments, and emergency repair preparations to be arranged in advance according to the risk arrival time of different devices. At the same time, this invention further uses information such as risk level, warning moment, and risk distribution to determine prevention and control actions such as equipment inspection, windproof reinforcement, operation adjustments, and emergency material deployment. It also uses actual faults and prevention and control results to update historical data and models, so that the risk prediction results can further serve proactive prevention and control, and form a closed loop of prediction, warning, prevention and control, and feedback updates. Attached Figure Description
[0025] Figure 1 This is a flowchart illustrating the overall method of an embodiment of the present invention; Figure 2 This is a schematic diagram of a typhoon disaster risk early warning system according to an embodiment of the present invention. Detailed Implementation
[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0027] Example 1: like Figure 1 As shown, this embodiment sequentially completes the fusion processing of multi-source power grid disaster data, the construction of a multi-dimensional disaster risk characterization system, the meteorological model-driven mapping of meteorological and power grid risks, the spatiotemporal dynamic evolution prediction of disaster risks, and the output of risk classification and prevention and control strategies. Each processing stage transmits data through equipment identifiers, unified timestamps, and equipment spatial locations, allowing data objects generated in the previous stage to directly serve as input for the next stage, thus avoiding a loss of correspondence between meteorological grids, equipment records, and risk results.
[0028] S1. Acquire typhoon meteorological data, power grid equipment data, spatial environment data, and historical fault data of the target power grid, and complete data association based on equipment identification and time to form equipment association data.
[0029] Typhoon meteorological data includes gridded meteorological forecast data and measured meteorological data. Gridded meteorological forecast data records precipitation, temperature, humidity, typhoon center location, typhoon intensity, typhoon path, maximum sustained wind speed, gust speed, wind direction, central pressure, and the radius of wind circles at different levels. Measured meteorological data comes from ground meteorological stations, weather radar, wind speed sensors, and rainfall monitoring devices. Power grid equipment data includes basic ledgers and operational status data for transmission and distribution equipment. Basic ledgers record at least equipment identification, equipment type, equipment location, line length, line direction, tower type, and years of operation. Operational status data records at least equipment health status, load level, maintenance status, and online monitoring data. Spatial environment data includes geographic information system data, digital elevation model data, land use types, river and road distribution, urban building distribution, underground space distribution, and drainage zoning data. Historical fault data includes equipment fault records, repair work orders, meteorological disaster process data, and disaster loss data corresponding to each typhoon, forming a historical fault sample according to the relationship between typhoon characteristics, equipment status, and fault results.
[0030] This embodiment uses the device identifier as the primary key for the device dimension and a Coordinated Universal Time (UTC) timestamp as the primary key for the time dimension. Meteorological forecast data is stored in a 1 km spatial grid with 15-minute time intervals. Device operating status data is accessed at the original sampling frequency and then resampled to 15-minute intervals. Spatial environment data is static or quasi-static data that changes slowly and is re-matched when device locations change or basic geographic data is updated. The fault start time, recovery time, and device identifier in historical fault records are indexed against the device's basic ledger. One device identifier corresponds to one device association record within a 15-minute time window. This record stores the device status vector, the spatial environment vector at the device location, and the historical fault label corresponding to that time window. Continuous device association records are arranged chronologically to provide a definite data order for the subsequent formation of a device risk characteristic sequence.
[0031] Before entering the correlation process, data integrity checks are performed. For meteorological time series data with no more than four consecutive time windows of missing data, linear interpolation is used. For typhoon center location and wind speed series with significant nonlinear changes, cubic spline interpolation is used. When gridded meteorological variables are missing, inverse distance-weighted spatial compensation is used, selecting the eight grids with the closest spatial distance and valid observations, with a distance power exponent of 2. When equipment operating status is missing for a short period, it is repaired based on the most recent valid status of the same equipment, the status of adjacent equipment with electrical connections, and the historical median operating value within the same load range. For data records with consecutive missing periods exceeding a preset maximum, a missing mask is added, and deterministic replacement values are not directly generated, thus enabling the model to distinguish between measured data and repaired data.
[0032] After missing data repair, each data sample is fed into an isolation forest for anomaly detection. The isolation forest consists of 200 random trees, each drawing 256 samples without replacement from the training data. The anomaly scoring formula is as follows: in, Represents data samples With a sample size of Abnormal scoring at that time; Represents data samples Path length in an isolated tree; This represents the expected value of the path length corresponding to multiple isolated trees; Indicates the sample size is The normalization factor for the average path length of the binary search tree. The 99th percentile of the anomaly score in the training data is used as the initial anomaly threshold; when the false positive rate on the continuous validation set exceeds 1%, the threshold is increased until the false positive rate meets the requirements. Data with anomaly scores exceeding the threshold is not immediately deleted, but undergoes secondary verification based on the operating status of adjacent devices, the upper limit of the change rate of similar sensors, and communication quality indicators. Data verified as sensor mutations are replaced with the median of adjacent time windows, and the anomaly mask is retained; data verified as real device mutations retain their original values.
[0033] After data cleaning, coordinate transformation, timestamp synchronization, unit conversion, and data format standardization are performed uniformly. Bilinear interpolation is used between meteorological grids and equipment locations; nearest neighbor matching is used for equipment located outside the grid boundaries and within one grid interval of the nearest valid grid; data records outside this range are marked as having no valid meteorological input. Continuous variables are standardized using the training set mean and standard deviation, while categorical variables are standardized using one-hot encoding or trainable embedding encoding. Standardized parameters are calculated only from the training set and directly used in the validation set, test set, and online inference data, avoiding future data participation in statistical calculations during the training phase.
[0034] Based on the cleaning and correlation results, four risk characteristics were constructed: meteorological impact, equipment vulnerability, spatial exposure, and historical disaster correlation. Meteorological impact characteristics include maximum sustained wind speed, gust speed, wind speed change rate, angle between wind direction and power line alignment, cumulative rainfall, maximum rainfall intensity, duration of continuous rainfall, soil moisture index, distance between equipment and typhoon center, wind circle coverage status, typhoon movement direction, and future impact duration. Equipment vulnerability characteristics include years of operation, number of historical failures, attributes of key power supply nodes, load level, and maintenance status. Equipment vulnerability is represented as: in, Indicates the vulnerability characteristics of the equipment; Represents a device vulnerability encoding function; This indicates the equipment's service life and aging information; This displays information on the number of historical failures and the types of failures. This indicates the equipment's health status, maintenance status, and reliability information; This indicates load level and operating condition information. The equipment vulnerability encoding function uses a two-layer fully connected network. The first layer has an output dimension of 64 and uses the ReLU activation function, while the second layer has an output dimension of 32. Another implementation uses gradient boosting trees trained separately for each equipment type to generate vulnerability scores, which is suitable for scenarios with a small number of historical samples and where the output basis is easily traceable.
[0035] Spatial exposure characteristics include topographic elevation, slope, distance from the river, urban building density, and underground space distribution, represented as: in, Indicates spatial exposure risk characteristics; Indicates the space exposure encoding function; Indicates terrain elevation, slope, and surface relief; This indicates land use types, river and road distribution, and underground space distribution; This indicates information on urban building density, drainage conditions, and other environmental exposures. It uses independent symbols. This represents terrain information to avoid confusion with subsequent power grid topology maps. Spatial exposure coding results remain unchanged at the device location and are recalculated after new digital elevation models, land use data, or drainage data are released.
[0036] Historical disaster association features are generated from a disaster knowledge graph. Graph entities include typhoon events, meteorological conditions, equipment, equipment type, failure mode, and spatial region. Relationships include at least impact, location, belonging, occurrence, and propagation. Equipment failure records and repair work orders are aligned into triples after entity alignment. A TransE relation embedding network is used to map entities and relationships into 64-dimensional vectors. The interval ranking loss is calculated using correct triples and negative triples formed by randomly replacing head or tail entities. For the current equipment, the one-hop entity directly connected to the equipment and the two-hop entity corresponding to the failure propagation relationship are taken, and average pooling is used to obtain historical disaster association features. When historical samples are limited, rule-based counting vectors are used instead of relation embedding. The counting vectors record the failure frequency of the same equipment type under similar wind speed ranges, rainfall ranges, and spatial environment types. This method has lower training overhead and can preserve the statistical interpretability of historical failure modes.
[0037] S2. Input typhoon meteorological data into the meteorological big data model to obtain the spatiotemporal characteristics of typhoon meteorology. Then, based on the equipment location and time, fuse the equipment-related data to form a sequence of equipment risk characteristics for each power grid device.
[0038] This embodiment employs a large-scale meteorological model based on a Transformer encoder. The model input consists of 24 consecutive 15-minute time intervals of meteorological grids, predicting the next 24 15-minute time intervals, corresponding to a 6-hour input window and a 6-hour prediction window. Each meteorological grid is first divided into 4×4 grid blocks. A two-dimensional convolutional embedding layer maps the meteorological variables within each grid block into a 256-dimensional meteorological feature vector, which is then added to the temporal location encoding, spatial grid location encoding, and meteorological variable type encoding. For each time interval... The original meteorological vector The embedding process is represented as: in, Indicates time Meteorological embedding vector; Represents the embedding weight matrix; Indicates time The original meteorological vector; This represents the embedding bias vector. Spatial grid position coding and temporal position coding are related to... Before addition, all are projected to 256 dimensions.
[0039] The large-scale meteorological model comprises six sequentially connected Transformer encoders. Each encoder includes a multi-head self-attention layer, a feedforward network, residual connections, and layer normalization. The multi-head self-attention layer has eight attention heads, each with a key vector dimension of 32. The feedforward network has a hidden dimension of 1024 and a dropout ratio of 0.1. A single attention head executes: in, Represents the query matrix; Represents the key matrix; Represents a value matrix; Indicates the dimension of the key vector; represents the transpose of the key matrix; Softmax represents the normalized exponential function. Different attention heads learn the correlations between meteorological variables, between adjacent time points, and between distant grids, respectively. The outputs of multiple heads are concatenated and then fed into the linear mapping layer.
[0040] The model output includes typhoon path, wind field, and rainfall outputs. The typhoon path output shows the typhoon's center location, speed, and direction of movement at future times; the wind field output shows the maximum sustained wind speed, gust speed, wind direction, and wind circle radius; and the rainfall output shows the rainfall intensity, cumulative rainfall, and duration of continuous rainfall. The results from the three outputs are combined according to the future time and spatial grid location to form the typhoon's spatiotemporal meteorological features. During the pre-training phase, multi-year historical meteorological data is used to complete mask reconstruction and multi-step prediction training. During the transfer adaptation phase, the first four Transformer encoders are frozen, and the last two Transformer encoders, layer normalization parameters, and three outputs are updated using historical typhoon data from the target power grid region. When there is sufficient data, all encoders are unfrozen and joint fine-tuning is performed using a smaller learning rate, which has the advantage of more thorough adaptation to local wind fields and rainfall patterns. When there are fewer samples, only the outputs are updated and the low-rank adaptation layers of the inserted encoders are updated, which has the advantage of reducing overfitting and computational resource requirements.
[0041] Equipment is extracted from the spatiotemporal characteristics of typhoon meteorology based on its location and future time. At any moment Corresponding meteorological characteristics Equipment data, spatial environment data, and historical fault data are respectively processed by a fully connected encoder, a spatial feature encoder, and a knowledge graph encoder to obtain equipment features. Spatial environment characteristics and historical fault characteristics All four feature classes are transformed into 128 dimensions through linear projection and layer normalization. The fused input is represented as: in, Indicates device At any moment The concatenated input features; Indicates meteorological characteristics; Indicates equipment characteristics; Indicates spatial environmental characteristics; Represents historical fault characteristics; square brackets indicate concatenation by feature dimension.
[0042] Equipment characteristics, spatial environment characteristics, and historical fault characteristics are spliced together and encoded into 128-dimensional power grid domain characteristics by a two-layer perceptron. The gated fusion network determines the gating coefficients based on four types of input features: in, Indicates device At any moment The corresponding 128-dimensional gating coefficient vector; This represents the Sigmoid activation function; Represents the gate weight matrix; This represents the gating bias vector; the meanings of the remaining symbols are the same as in the previous formula. Each element of the gating coefficient is between 0 and 1, used to allocate the proportion of meteorological information and power grid information on the same feature dimension.
[0043] equipment At any moment The equipment risk characteristics are represented as follows: in, Indicates device At any moment The equipment risk characteristics; Indicates characteristics of the power grid sector; This indicates element-wise multiplication; This represents the power grid sector gating coefficient, which is complementary to the meteorological gating coefficient. If the original dimensions of the meteorological features and the power grid sector features are different, they are projected to the same dimension using a linear layer before performing element-wise multiplication.
[0044] Under the general expression that omits device and time indexes, the above fusion relationship can be represented as: in, Indicates general fusion features; Indicates the general gating coefficient; Indicates general meteorological characteristics; Indicates characteristics of the general power grid sector; This indicates that corresponding elements are multiplied. and Both express the same gating logic; the former describes the case without an index, while the latter is used for device-level timing calculations. They represent the corresponding time intervals of the same device at multiple consecutive moments. Arranged chronologically, this forms a sequence of equipment risk characteristics for the device.
[0045] S3. Construct a power grid topology map based on the power grid equipment and the electrical connections between the equipment. Input the equipment risk feature sequence into a graph neural network risk prediction model based on the power grid topology map to obtain the future risk sequence of each equipment.
[0046] The power grid topology is represented as follows: in, Represents the power grid topology; Represents a set of device nodes; This represents a set of electrical connection edges. Substations, towers, line sections, distribution transformers, and switchgear are respectively designated as equipment nodes. Physical line connections, busbar connections, and switch conduction relationships form electrical connection edges. When the switch state changes, the effective state of the corresponding edge is updated to ensure the online topology is consistent with the current operating mode. Directed edges are used for directional power supply relationships, while for scenarios where only mutual influence needs to be expressed, two edges in opposite directions are combined into a bidirectional connection.
[0047] The device risk feature sequence of each device node first enters a bidirectional gated recurrent unit for time-dimensional encoding, with a hidden dimension of 128. The forward and reverse states are concatenated and linearly projected onto the 128-dimensional array to obtain the initial node state at each time step. When online real-time inference does not allow the use of future measured values, a unidirectional gated recurrent unit is used to ensure the time-dimensional representation is accurate. The state is only determined by And the previous data formation. After completing time encoding, the graph neural network transmits and aggregates node states between adjacent device nodes, the 1st... The layer node state is updated according to the following formula: in, Represents device node At any moment The Layer node status; Represents device node At any moment The Layer node status; Indicates adjacent device nodes At any moment The Layer node status; Indicates the relationship with device nodes A set of adjacent nodes that have valid electrical connections; This represents the weight matrix corresponding to the node's own state. The weight matrix represents the state of adjacent nodes; This represents the ReLU nonlinear activation function. To reduce aggregation scaling bias caused by differences in node degree, the summation result of adjacent nodes is divided by in actual calculations. Alternatively, the normalization coefficient can be incorporated into the edge weights.
[0048] This embodiment employs two layers of graph convolution to update node states, followed by two layers of unidirectional gated recurrent units as a multi-step temporal decoder. The decoder has a hidden dimension of 128. The first future time step uses the graph convolution output and the meteorological features of the last measured time step as input. Subsequent time steps use the hidden state of the previous time step and the corresponding future meteorological features as input. During the training phase, the true previous time step label is used with an initial probability of 0.5, and the probability linearly decreases to 0 with each training round. During the online inference phase, only the prediction results of the previous time step are used. The decoder output sequentially feeds into three linear output heads: fault probability, fault time, and propagation location, forming the equipment fault probability for the next 24 time steps and auxiliary training results. In the future The failure probability is expressed as: in, Indicates device In the future The probability of an internal failure; This represents a probabilistic prediction function composed of a time encoder, a graph neural network, and a time decoder; Indicates device At any moment The fusion characteristics; Indicated by device node The local power grid topology subgraph centered on [the grid]. Actual calculations use the aforementioned [method / approach]. As .
[0049] The importance of equipment is determined based on the power supply user level, load size, attributes of important power nodes, and the scope of fault impact. The four indicators are normalized separately and then weighted and summed. The weights are initially determined by operation management experts using the analytic hierarchy process (AHP), and then calibrated using historical power outage loss data. When power outage loss data is unavailable, all weights are equal. The equipment risk index is: in, Indicates device In the future The equipment risk index; Indicates the probability of equipment failure; Indicates the importance of equipment. It compares the importance of the same equipment over 24 consecutive future time points. Arranged chronologically, these form a sequence of future risks.
[0050] The model is trained under supervised supervision using historical typhoon events. The training sample set is represented as follows: in, Represents the training sample set; Indicates the first Each sample's equipment risk characteristic sequence and corresponding power grid topology; Indicates the first Each sample is labeled with its corresponding equipment fault status, fault occurrence time, and fault propagation location. This indicates the number of training samples. The samples are divided into training, validation, and test sets according to typhoon events. Data from the same typhoon event can only appear in one of the sets to prevent data from leaking into different sets from adjacent time windows.
[0051] The joint loss function is expressed as: in, Indicates joint loss; The weighted binary cross-entropy loss represents the classification of equipment faults. This represents Huber's losses at the moment the failure occurred; Cross-entropy loss representing the location of fault propagation; , and These represent the weights of the three loss terms. The initial weights are set to 1, 0.5, and 0.5, and then normalized based on the magnitudes of the three losses on the validation set, ensuring that the average gradient norm of any loss term does not exceed twice the average gradient norm of the other loss terms. When there is a positive-negative sample imbalance, The positive sample weights are taken as the ratio of the number of negative samples to the number of positive samples in the training set.
[0052] To describe the dynamic coupling between typhoon path and equipment location, equipment data is obtained. Location And the future Typhoon center location Calculate the spatial distance between the two: in, Indicates device Location; and These represent the horizontal and vertical coordinates of the device's location, respectively. Indicates future time The location of the typhoon's center; and These represent the horizontal and vertical coordinates of the typhoon's center location, respectively. This indicates the spatial distance between the equipment and the typhoon center. The coordinates are projected coordinates in meters; if the input is latitude and longitude, it is first converted to the target power grid's unified projected coordinate system before the distance is calculated.
[0053] Combining future moments The radius of the typhoon's wind circle determines the degree of its direct impact: in, Indicates device In the future The extent of the direct impact of the typhoon; Indicates future time The radius of the typhoon's wind circle; The meaning is the same as the previous formula; This represents a natural exponential function. When there are wind circles of level 7, 10, and 12, the degree of direct impact is calculated respectively, and weighted according to the corresponding wind circle level weights, with the weights increasing as the wind circle level increases.
[0054] Further combining wind speed, wind direction, the angle between the line and the equipment location, and the terrain at the equipment location to form a comprehensive meteorological influencing factor: in, Indicates device In the future Comprehensive meteorological influencing factors; Indicates the degree of direct impact of the typhoon; This represents the wind speed influence coefficient, determined by wind speed and the wind resistance rating of the equipment. This represents the directional influence coefficient determined by the angle between the wind direction and the route alignment; This represents the topographic influence coefficient determined by elevation, slope, and surface roughness. Each influence coefficient is mapped to 0.5 to 1.5 by the training set quantiles, and the product is truncated and normalized to 0 to 1. For risk correction, the equipment risk index and the comprehensive meteorological influence factor are assigned weights of 0.7 and 0.3 respectively, summed, and then truncated to 0 to 1. These weights are determined through a grid search on the validation set, and the sum of the two weights remains 1. This process limits the amplification of the final risk by a single meteorological factor.
[0055] To express continuous risk states, time will be... All equipment risks can be represented as a risk state matrix: in, Indicates time The power grid risk state matrix; They represent the first The device node to the first Each device node at time The correction risk value; Indicates the number of device nodes. Based on the current risk status and meteorological input, the future [number]th [node] is predicted. Risk status over time intervals: in, Indicates future time The risk state matrix; This represents the risk evolution function consisting of temporal attention, graph propagation, and temporal decoding. Indicates time Corresponding typhoon meteorological spatiotemporal characteristics and equipment status characteristics; Indicates the sequence number of the future time interval.
[0056] The risk propagation strength of electrical connection edges is determined based on edge impedance, line length, and switch status. The edge weight mapping function employs a two-layer perceptron with an input dimension of 3, a hidden dimension of 16, and an output dimension of 1. The hidden layer uses ReLU activation, and the output is normalized using Softmax within the same set of neighboring nodes, as follows: in, Represents device node To device node Risk propagation weight; This represents the edge weight mapping function; Represents a node With nodes The impedance of the connecting lines; Indicates the length of the connecting line; This indicates the on / off state of the switch. Edge weights are normalized within the set of adjacent nodes; when the switch is off, the corresponding edge weights are reset to 0.
[0057] Risk diffusion update based on edge weights: in, Represents device node The risk status of transmission in the next moment; Represents device node The current risk situation; Represents a node To the node Risk propagation weight; This represents the risk propagation weight matrix; Represents a nonlinear activation function; the summation range is the summation interval between nodes. There are nodes with valid electrical connections. This formula is used to explain the propagation process of risk in the power grid topology and is trained together with the aforementioned graph neural network node state updates.
[0058] S4. Determine the risk level and warning time of each device based on the future risk sequence, and generate typhoon disaster risk warning information.
[0059] This embodiment determines the risk level threshold on the validation set based on risk recall rate, false positive rate, and operational resource capacity. Example thresholds are 0.25, 0.50, and 0.75: a risk index less than 0.25 indicates low risk, 0.25 and less than 0.50 indicates medium risk, 0.50 and less than 0.75 indicates high risk, and 0.75 indicates extremely high risk. When deployed to different voltage levels or equipment types, the thresholds are re-determined using corresponding validation data. The warning threshold is set to the lower limit of high risk, 0.50. The future moment when the corrected risk index of a device first reaches 0.50 is recorded as... The current time is recorded as The advance warning time is expressed as: in, Indicates the lead time for early warning; This indicates the future moment when the risk index first reaches the warning threshold; This indicates the current moment when the warning is generated. If none of the risk indices within the future forecast window reach the warning threshold, no device-level high-risk warning will be generated, but the complete future risk trend will be retained.
[0060] The system generates risk area distribution data, a list of high-risk equipment, estimated failure time data, equipment risk ranking data, and future risk trend data based on future risk sequences, risk levels, and warning times. To balance equipment failure probability, equipment importance, and load impact, a comprehensive risk score can also be used. in, Indicates device Comprehensive risk score; Indicates the maximum probability of device failure within a specified future window; Indicates the importance of the equipment; Indicates the degree of influence of normalized load; , and The three weights represent non-negative scoring weights, and their sum is 1. In this embodiment, the three weights are set to 0.5, 0.3, and 0.2 respectively; when sufficient historical power outage loss data is available, logistic regression is used to calibrate the three weights on the validation set.
[0061] The input vector of the prevention and control strategy recommendation model is represented as: in, This represents the input vector for the prevention and control strategy; Indicates device The maximum risk index within the warning window; Indicates the lead time for early warning; This represents a device attribute vector, which includes at least device type, device location, health status, maintenance status, load level, and key power supply node attributes. The candidate set of control actions is represented as: in, Indicates device The corresponding set of candidate prevention and control actions; to These respectively represent equipment inspection, windproofing reinforcement, operation mode adjustment, emergency material deployment, and other actions permitted by safety rules; This indicates the number of candidate actions. Before execution, actions are verified according to power grid safety rules. Actions that do not meet power flow, equipment operation permission, or personnel safety constraints are removed from the candidate set.
[0062] When using reinforcement learning to optimize candidate prevention and control actions, the states include the current typhoon status, power grid operation status, and risk distribution status. The actions correspond to candidate prevention and control measures validated by safety rules. For discrete action scenarios, Double DQN is used. Both the online network and the target network contain two fully connected layers with a hidden dimension of 128. The experience replay pool sets sampling priorities according to the actual changes in disaster losses, and the target network updates its parameters every 500 iterations. The estimated disaster loss is calculated using a gradient boosting regression tree trained on historical fault durations, affected loads, and user levels. The action execution cost is derived from the operation and maintenance resource configuration table. The safety status score is determined after normalizing power flow exceedance, load loss, and power supply status of important nodes. The action evaluation function is: in, Indicates the evaluation value of the candidate prevention and control action; Indicates the expected disaster loss after the action is taken; Indicates the cost of performing the action; This indicates the power grid safety status score after the action is performed; , and These represent the evaluation weights for disaster loss, execution costs, and safety status, respectively. The candidate action with the highest evaluation value is written into the early warning information.
[0063] Early warning information is sent to the power grid dispatch platform, emergency command platform, and mobile terminals. The transmission record includes equipment identification, risk level, warning time, warning lead time, risk trend, risk influencing factors, and target prevention and control actions. After the typhoon, the actual equipment failure status, actual failure time, failure propagation location, and actual prevention and control results are written back to the historical failure data. New data, after quality verification, enters the incremental training pool; when the calibration error of the equipment failure probability of the most recent typhoon event exceeds 20% of the historical baseline, or the cumulative number of newly added valid samples exceeds 10% of the original training set, the graph neural network risk prediction model is updated. The prevention and control strategy model adjusts the action evaluation parameters according to actual losses, actual costs, and safety status, thus forming a closed loop of data collection, risk prediction, proactive defense, disaster feedback, and model optimization.
[0064] Example 2: This embodiment provides a typhoon disaster risk early warning system based on fused data and a large meteorological model, such as... Figure 2 As shown, it includes: The data acquisition module is used to acquire typhoon meteorological data, power grid equipment data, spatial environment data, and historical fault data of the target power grid. The typhoon meteorological data includes meteorological information such as typhoon path, typhoon intensity, wind speed, and precipitation. The power grid equipment data includes equipment location, equipment type, operating status, and online monitoring data. The spatial environment data includes geographic information, terrain features, and environmental exposure information. The historical fault data includes equipment fault records and disaster loss data during historical typhoon disasters. The data association module is used to associate the power grid equipment data, spatial environment data and historical fault data based on the equipment identification, equipment location and time information, and to perform data verification, missing data repair, anomaly detection and format standardization processing on multi-source data to form equipment association data; The meteorological feature extraction module is used to input the typhoon meteorological data into the meteorological big model, and extract the typhoon path features, wind field diffusion features and rainfall evolution features by performing correlation analysis on meteorological variables at different times and spatial locations, so as to obtain the typhoon meteorological spatiotemporal features. The risk feature fusion module is used to match the typhoon meteorological spatiotemporal features with the equipment-related data based on the equipment location and time, and to fuse meteorological features, equipment features, spatial environment features and historical disaster-related features to generate a sequence of equipment risk features for each power grid device. The risk prediction module is used to construct a power grid topology map based on the electrical connection relationship between power grid equipment, and input the risk feature sequence of the equipment into a graph neural network risk prediction model based on the power grid topology map. Through information transmission and feature aggregation between equipment nodes, the future risk sequence of each power grid equipment at multiple future times is obtained. The risk warning generation module is used to determine the risk level and warning time corresponding to each power grid device based on the future risk sequence, and generate typhoon disaster risk warning information that includes risk area distribution, high-risk device list, expected failure time and risk trend information; The prevention and control strategy recommendation module is used to determine corresponding prevention and control measures such as equipment inspection, windproof reinforcement, operation mode adjustment and emergency material deployment based on the risk level, early warning time, power grid operation status and risk distribution status, and to write the prevention and control measures into the typhoon disaster risk early warning information.
[0065] The typhoon disaster risk early warning system provided in this embodiment achieves the fusion analysis of meteorological information, power grid equipment status, spatial environmental factors, and historical disaster information through the collaborative work of the data acquisition module, data association module, meteorological feature extraction module, risk feature fusion module, risk prediction module, and risk early warning generation module. It can dynamically predict the future risks of power grid equipment under the influence of typhoon disasters and further output equipment-level risk early warning results and prevention and control suggestions, thereby improving the power grid's disaster risk perception and proactive defense capabilities under extreme typhoon conditions.
[0066] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0067] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A typhoon disaster risk early warning method based on fusion data and a weather macro model, characterized in that, Includes the following steps: Acquire typhoon meteorological data, power grid equipment data, spatial environment data, and historical fault data of the target power grid; and correlate the power grid equipment data, spatial environment data, and historical fault data based on equipment identification and time to obtain equipment correlation data. The typhoon meteorological data is input into the meteorological big model to obtain the typhoon meteorological spatiotemporal characteristics. The typhoon meteorological spatiotemporal characteristics are then fused with the equipment-related data based on the equipment location and time to obtain the equipment risk characteristic sequence of each power grid equipment. A power grid topology map is constructed based on the power grid equipment and the electrical connections between the power grid equipment. The risk feature sequence of the equipment is input into a graph neural network risk prediction model based on the power grid topology map to obtain the future risk sequence of each power grid equipment. Based on the future risk sequence, the risk level and warning time of each power grid device are determined, and typhoon disaster risk warning information is generated. 2.The typhoon disaster risk early warning method based on fusion data and weather general model according to claim 1, wherein, The acquisition of typhoon meteorological data, power grid equipment data, spatial environment data, and historical fault data of the target power grid specifically includes: Gridded meteorological forecast data and measured meteorological data, including precipitation, temperature, humidity, typhoon center location, typhoon intensity, typhoon movement path and wind speed, are obtained as the typhoon meteorological data; Acquire basic ledgers and operating status data of power transmission and distribution equipment, including equipment location, line length, tower type, equipment commissioning years, equipment health status, load level, and online monitoring data, as the power grid equipment data; Obtain geospatial data including geographic information system data, digital elevation model data, land use types, river and road distribution, and urban building distribution, as the spatial environment data; Acquire equipment failure records, emergency repair work orders, meteorological disaster process data, and disaster loss data corresponding to each typhoon disaster, and form historical failure samples according to the correspondence between typhoon characteristics, equipment status, and failure results, as the historical failure data. 3.The typhoon disaster risk early warning method based on fusion data and weather general model according to claim 1, characterized in that, The process of associating power grid equipment data, spatial environment data, and historical fault data based on equipment identification and time to obtain equipment association data specifically includes: The integrity of the typhoon meteorological data, power grid equipment data, spatial environment data and historical fault data is checked. Time interpolation is performed on the missing meteorological time series, and inverse distance weighted spatial compensation is performed on the missing gridded meteorological data. The missing equipment operating status is repaired by combining the electrical connection relationship between power grid equipment and historical operating data. The missing data is input into the isolated forest model to obtain the anomaly score of each data sample. Data samples with anomaly scores exceeding the preset anomaly threshold are identified as anomalous data. The anomalous data is then verified based on the operating status of the adjacent equipment of the power grid equipment corresponding to the anomalous data. The data that has completed the abnormal data verification is subjected to coordinate transformation, timestamp synchronization and data format standardization. The spatial environment data is spatially matched with the corresponding power grid equipment data according to the equipment location. The historical fault data is associated with the corresponding power grid equipment data according to the equipment identifier and timestamp. Data on power grid equipment, spatial environment, and historical faults that have the same device identifier and are located within the same time window are combined into a single device association record. These device association records are then arranged in chronological order to obtain the device association data. 4.The typhoon disaster risk early warning method based on fusion data and weather general model according to claim 1, wherein, The process of fusing the typhoon meteorological spatiotemporal characteristics with the associated equipment data based on equipment location and time to obtain an equipment risk characteristic sequence for each power grid device specifically includes: From the aforementioned typhoon meteorological spatiotemporal characteristics, wind field characteristics, rainfall and urban flooding characteristics, and typhoon spatial impact characteristics are extracted. The wind field characteristics include maximum sustained wind speed, gust speed, wind speed change rate, and the angle between wind direction and line direction. The rainfall and urban flooding characteristics include cumulative rainfall, maximum rainfall intensity, continuous rainfall duration, and soil moisture index. The typhoon spatial impact characteristics include the distance between the typhoon center and the equipment location, the wind circle coverage status, the typhoon's movement direction, and the future impact time. The equipment vulnerability features and spatial exposure features are extracted from the equipment association data. The equipment vulnerability features include the equipment's years of operation, number of historical failures, attributes of important power supply nodes, load level, and maintenance status. The spatial exposure features include terrain elevation, slope, distance from the river, urban building density, and underground space distribution. Based on the historical fault data, a disaster knowledge graph containing the relationships between typhoon type, meteorological conditions, equipment type and fault mode is constructed, and historical disaster association features are extracted from the disaster knowledge graph based on the entities and relationships corresponding to each power grid equipment. Based on the location and time of the equipment, the wind field characteristics, rainfall and waterlogging characteristics, typhoon spatial impact characteristics, equipment vulnerability characteristics, spatial exposure characteristics, and historical disaster association characteristics are matched with the corresponding power grid equipment. The characteristics of the same power grid equipment at multiple consecutive times are arranged in chronological order to obtain the equipment risk characteristic sequence. 5.The typhoon disaster risk early warning method based on fusion data and weather general model according to claim 1, wherein, The step of inputting the typhoon meteorological data into a large meteorological model to obtain the spatiotemporal characteristics of typhoon meteorology specifically includes: Typhoon meteorological data from multiple consecutive moments are combined into a meteorological input sequence in chronological order. The meteorological variables, temporal location, and spatial grid location in the meteorological input sequence are encoded into meteorological feature vectors through an embedding layer. The meteorological feature vector is input into a large meteorological model based on the Transformer architecture. The correlation features between different meteorological variables, different time locations, and different spatial grid locations are extracted through a multi-head self-attention layer. The correlation features are then iteratively encoded through multiple Transformer encoders connected in sequence. The output layer of the meteorological model outputs the typhoon path characteristics, wind field diffusion characteristics, and rainfall evolution characteristics corresponding to multiple future time points, and combines the typhoon path characteristics, wind field diffusion characteristics, and rainfall evolution characteristics according to time and spatial grid positions to obtain the typhoon meteorological spatiotemporal characteristics. The meteorological model is pre-trained using historical meteorological data and adapted using historical typhoon meteorological data of the target power grid area. 6.The typhoon disaster risk early warning method based on fusion data and weather general model according to claim 5, wherein, The process of fusing the typhoon meteorological spatiotemporal characteristics with the associated equipment data based on equipment location and time to obtain an equipment risk characteristic sequence for each power grid device specifically includes: Based on the equipment location and time, extract the relevant information about the power grid equipment from the spatiotemporal characteristics of the typhoon meteorological data. At any moment Corresponding meteorological characteristics The device features are obtained by feature encoding the power grid equipment data, spatial environment data, and historical fault data in the device-related data. Spatial environment characteristics and historical fault characteristics ; The device features Spatial environment characteristics and historical fault characteristics Concatenate and encode into power grid features And determine the gating coefficient: Regarding the meteorological characteristics and the characteristics of the power grid field Perform gating fusion to obtain equipment At any moment Corresponding equipment risk characteristics: in, Indicates device At any moment The corresponding gating coefficient; This represents the Sigmoid activation function; Represents the gate weight matrix; Represents the gate bias vector; square brackets indicate feature concatenation; Indicates meteorological characteristics; Indicates equipment characteristics; Indicates spatial environmental characteristics; Indicates historical fault characteristics; Indicates characteristics of the power grid sector; Indicates the risk characteristics of the equipment; This indicates element-wise multiplication; The equipment risk characteristics of the same power grid equipment at multiple consecutive times Arranged in chronological order, the sequence of risk characteristics of the equipment is obtained.
7. The typhoon disaster risk early warning method based on fused data and a meteorological big data model according to claim 1, characterized in that, The process of constructing a power grid topology map based on the power grid equipment and the electrical connections between the equipment, and inputting the equipment risk feature sequences into a graph neural network risk prediction model based on the power grid topology map to obtain the future risk sequences of each power grid equipment, specifically includes: The power grid equipment is set as a node in the power grid topology graph, the electrical connection relationship between different power grid equipment is set as an edge in the power grid topology graph, and the node characteristics of the corresponding node are set according to the equipment risk characteristic sequence corresponding to each power grid equipment. The device risk feature sequences of each node are encoded in a time dimension to obtain the initial node state of each node at different times. The node states are then transmitted and aggregated between adjacent nodes using the graph neural network risk prediction model. Wherein, the first... The node state of a layer is defined by the following formula: in, Represents device node At any moment The corresponding number Layer node status; Represents device node At any moment The corresponding number Layer node status; Indicates adjacent device nodes At any moment The corresponding number Layer node status; Indicates the relationship with device nodes A set of adjacent device nodes that are electrically connected; This represents the weight matrix corresponding to the node's own state. The weight matrix represents the state of adjacent nodes; Represents a non-linear activation function; The status of each device node that has completed the node status update is decoded according to future time points to obtain the future risk sequence of each power grid device at multiple consecutive future time points.
8. The typhoon disaster risk early warning method based on fused data and a meteorological big data model according to claim 1, characterized in that, The graph neural network risk prediction model is trained under supervised supervision using the historical fault data, specifically including: The equipment risk feature sequence and power grid topology map corresponding to each power grid device during historical typhoons are used as training inputs. The corresponding power grid devices are used as training labels for whether they have failed, the time of failure, and the location of failure propagation during the historical typhoons. Risk classification loss, time prediction loss, and spatial prediction loss are used to jointly optimize the graph neural network risk prediction model. The device that outputs the trained graph neural network risk prediction model In the future The probability of an internal failure is defined as the equipment failure probability. The equipment is determined based on the power supply user level, load size, attributes of important power nodes, and the scope of fault impact. Equipment importance ; Determine equipment In the future Corresponding equipment risk index: in, Indicates device In the future The corresponding equipment risk index; Indicates device In the future The probability of equipment failure occurring within the system; Indicates device The importance of the equipment; The future risk sequence is formed by arranging the equipment risk indices corresponding to the same power grid equipment at multiple consecutive future moments in chronological order.
9. A typhoon disaster risk early warning method based on fused data and a meteorological big data model according to claim 8, characterized in that, The step of determining the risk level and early warning time of each power grid device based on the future risk sequence specifically includes: Acquire equipment Equipment location And the future Corresponding typhoon center location Determine the equipment Spatial distance from the center of the typhoon: Based on the spatial distance and future time The corresponding typhoon wind circle radius determines the equipment. In the future The corresponding direct impact of the typhoon: in, Indicates device The location of the equipment; Representing the equipment The horizontal and vertical coordinates; Indicates future time The corresponding location of the typhoon's center; These represent the horizontal and vertical coordinates of the typhoon's center location, respectively. Indicates device Spatial distance from the center of the typhoon; Indicates future time The corresponding radius of the typhoon's wind circle; Indicates device In the future The corresponding degree of direct impact of the typhoon; Based on the degree of direct impact of the typhoon, the angle between the wind direction and the route, and the terrain features corresponding to the equipment location, the equipment risk index in the future risk sequence is corrected to obtain the corrected future risk sequence. The equipment risk index in the corrected future risk sequence is compared with the preset risk level threshold to determine the risk level of each power grid equipment at different future times. The future time when the equipment risk index first reaches the warning threshold is determined as the warning time of the corresponding power grid equipment. The time difference between the warning time and the current time is determined as the warning lead time.
10. The typhoon disaster risk early warning method based on fused data and a meteorological big data model according to claim 1, characterized in that, The generation of typhoon disaster risk warning information specifically includes: Based on the future risk sequence, risk level and warning time of each power grid equipment, generate risk area distribution data, high-risk equipment list, expected failure time data, equipment risk ranking data and future risk trend data; The risk level, warning time, power grid equipment data, and risk area distribution data are input into the prevention and control strategy recommendation model to construct a state vector containing the current typhoon status, power grid operation status, and risk distribution status. Equipment inspection, windproof reinforcement, operation mode adjustment, and emergency material deployment are set as candidate prevention and control actions. The action evaluation value is determined based on the expected disaster loss, execution cost, and power grid safety status corresponding to the candidate prevention and control actions. Based on the action evaluation value, the target prevention and control action corresponding to each high-risk equipment is determined from the candidate prevention and control actions, and the target prevention and control action is written into the typhoon disaster risk early warning information; The typhoon disaster risk warning information is sent to the power grid dispatching platform, emergency command platform and mobile terminal. The actual equipment failure data, actual failure time and actual prevention and control results after the typhoon disaster are written into the historical failure data. The model parameters of the graph neural network risk prediction model and the prevention and control strategy data of the prevention and control strategy recommendation model are updated.
Citation Information
Patent Citations
Power grid risk early warning method and system under typhoon disaster based on multi-source data fusion
CN115688032A
