Fire risk level prediction system for power transmission channel based on feature fusion

Abstract

This invention relates to the fields of artificial intelligence and power safety monitoring technology, specifically a feature fusion-based fire risk level prediction system for power transmission channels. It aims to address the problems of spatial homogenization and response lag in risk assessment caused by the single feature representation, shallow fusion mechanism, and insufficient spatiotemporal modeling in existing technologies. The system comprises a data acquisition layer, a feature engineering layer, a fusion modeling layer, and a risk output layer. By incorporating five types of data—remote sensing, meteorology, line monitoring, terrain, and historical fire points—it constructs an 18-dimensional feature vector with 100-meter-level grid cells. A graph attention network and spatiotemporal convolution hybrid model are used to achieve cross-modal adaptive fusion, ultimately outputting a minute-level dynamic fire risk level map. This system supports automatic triggering of graded early warnings, significantly improving the accuracy and timeliness of wildfire risk prediction for power transmission channels.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and power safety monitoring technology, and in particular to a fire risk level prediction system for power transmission channels based on feature fusion. Background Technology

[0002] With the continuous expansion of power systems and the increasing prevalence of transmission lines traversing complex terrains (such as forests and mountains), the risk of fires caused by external environmental factors (such as forest fires, hot and dry climates, and contact with vegetation) along transmission lines has significantly increased. To ensure the safe operation of the power grid, it is urgent to conduct accurate, dynamic, and graded predictions of fire risks along transmission lines. Although existing technologies have conducted fire risk assessments in specific scenarios, when applied to linear infrastructure such as transmission lines, they still suffer from problems such as limited feature dimensions, insufficient fusion mechanisms, and coarse prediction granularity, making it difficult to support refined prevention and control decisions.

[0003] A search revealed a calculation method for superimposing algorithms on power equipment crossing forest areas with fire risk, published under publication number CN118735273B. This patent addresses the scenario of power equipment crossing forest areas by constructing risk functions for forest areas with different risk levels and performing superimposed calculations to achieve an overall risk assessment of the crossing area. However, this scheme is based solely on the risk level of the forest area and the probability of triggering factors, failing to fully integrate multi-source heterogeneous features such as meteorology, vegetation, topography, and line operation status. Furthermore, its selection of "featured fire risk forest areas" relies on the lowest risk level area, which is logically flawed and may lead to an underestimation of high-risk areas. In addition, its risk functions are statically superimposed, lacking the ability to model spatiotemporally dynamic changes, making it difficult to reflect the real-time evolution trend of fire risk along the transmission line.

[0004] A search revealed a method and system for fire risk analysis of cable trays, published under publication number CN116663910B. This patent focuses on the pre-analysis and risk projection of fire hazards inside cable trays, achieving risk level assessment through basic data collection and fire source environment analysis. Although it emphasizes pre-prediction, its application is limited to enclosed cable trays, failing to consider the complex external fire sources (such as wildfires and man-made fires) faced by power transmission channels in open environments. Furthermore, this method primarily relies on structured basic data, without incorporating multimodal data such as remote sensing, meteorological, and video surveillance, and does not employ feature fusion mechanisms for collaborative modeling of information from different sources. This limits its applicability to fire risk prediction in wide-area, open, and dynamic power transmission channels.

[0005] The aforementioned problems indicate that existing technologies for predicting fire risk levels in power transmission channels generally suffer from drawbacks such as limited feature dimensions, lack of fusion mechanisms, weak dynamic response capabilities, and insufficient scenario adaptability. Therefore, this invention proposes a feature fusion-based power transmission channel fire risk level prediction system. This system aims to construct a dynamic risk assessment model with spatiotemporal awareness through deep fusion of multi-source heterogeneous features (including geographic information, meteorological data, remote sensing imagery, line operating parameters, historical fire points, etc.). This enables high-precision, fine-grained, and real-time prediction of fire risk levels across the entire power transmission channel, providing scientific decision support for power grid disaster prevention and mitigation. Summary of the Invention

[0006] This invention provides a feature fusion-based fire risk level prediction system for power transmission channels, aiming to address the technical shortcomings of existing technologies, such as limited feature representation dimensions, shallow fusion mechanisms, and insufficient spatiotemporal dynamic modeling capabilities, leading to spatial homogenization and delayed response in risk assessment results. To achieve the above-mentioned objectives, this invention constructs a multi-level, multi-modal, high-dimensional heterogeneous data collaboratively driven dynamic fire risk prediction architecture. Its core lies in the organic integration of a geospatial semantic coding module, a meteorological-vegetation coupled state perception module, a line operation condition monitoring module, a historical fire behavior pattern mining module, and a cross-modal feature adaptive fusion prediction module, enabling minute-level dynamic prediction of fire risk levels within 100-meter-level spatial units along the power transmission channel.

[0007] The system comprises four logical layers: a data acquisition layer, a feature engineering layer, a fusion modeling layer, and a risk output layer. The data acquisition layer is responsible for acquiring raw observation data in real time from multi-source heterogeneous sensors and information systems; the feature engineering layer performs spatiotemporal alignment, physical quantity normalization, semantic enhancement, and feature vectorization on the raw data; the fusion modeling layer constructs nonlinear coupling relationships between cross-modal features based on deep neural networks and graph structure modeling techniques; and the risk output layer generates a standardized fire risk level map based on the output of the fusion model and supports the automatic triggering of graded early warning commands.

[0008] At the data acquisition layer, the system accesses five core data sources: The first type is high-resolution remote sensing image data, which comes from multispectral imagers mounted on low-orbit satellites or relay UAV platforms. Its spatial resolution is 0.5 meters to 2 meters, the time revisit period is no more than 6 hours, and the spectral bands cover visible light (450–700 nm), near-infrared (700–1000 nm), and short-wave infrared (1000–2500 nm), used to extract land cover type, vegetation index (NDVI), fire sensitivity index (BAI), and land surface temperature (LST); The second type is numerical weather prediction data, with a spatial grid accuracy of 1 km × 1 km and a time step of 1 hour, including wind speed (m / s), wind direction (°), relative humidity (%), air temperature (°C), precipitation (mm / h), and fuel dryness index (FFMC). The third category is online monitoring data for transmission lines, sourced from distributed fiber optic temperature measurement units, current transformers, and video monitoring terminals deployed on towers. The fiber optic temperature measurement units have a spatial sampling interval of 1 meter and a temperature measurement accuracy of ±0.5℃. The video monitoring terminals have a frame rate of 25fps and support the extraction of visual features of smoke and open flames. The fourth category is digital elevation model (DEM) and land use data, sourced from the 1:50,000 scale geographic database released by the National Geomatics Center of China. The horizontal positioning error is less than 5 meters, and the data is used to calculate slope (°), aspect (°), and terrain shading factor. The fifth category is historical fire point record data, sourced from the fire monitoring platform of the National Forestry and Grassland Administration. This data includes verified fire point coordinates, ignition time, burned area, and fire intensity level over the past five years. The timestamp accuracy is at the minute level, and the spatial positioning error is less than 30 meters.

[0009] At the feature engineering layer, the system performs a unified spatiotemporal benchmark alignment operation on the five types of raw data mentioned above. All spatial data are projected to the WGS84 geographic coordinate system and resampled to a unified 100m × 100m regular grid cell, with each grid cell corresponding to a basic risk assessment unit along the transmission channel. In the time dimension, all data streams are aligned to a 10-minute time slice using UTC timestamps as the reference, through linear interpolation or nearest neighbor padding. Based on this, each data source is transformed into a feature vector with clear physical meaning:

[0010] For remote sensing image data, the system calculates the normalized vegetation index within each grid cell. ,in and The surface reflectance is calculated in the near-infrared and red light bands, respectively; the combustion area index is also calculated. Surface temperature Obtained through inversion using a single-window algorithm, the formula is as follows: ,in This refers to the radiance in the thermal infrared band. , .

[0011] For meteorological data, the system directly extracts the WRF output value corresponding to each grid cell and further calculates the corrected dryness index of combustibles. ,in The relative humidity gradient reflects the abrupt changes in local microclimate.

[0012] Based on the line monitoring data, the system will use fiber optic temperature measurement sequences. The tower locations are mapped to corresponding grid cells, and the average conductor temperature within each cell is calculated. ,in The number of temperature measurement points falling into the grid; video surveillance data is used to extract smoke confidence scores through a pre-trained convolutional neural network (ResNet-18 architecture). The network was pre-trained on ImageNet and then fine-tuned using a power scene dataset containing 100,000 labeled smoke / non-smoke images. The loss function was cross-entropy, the optimizer was Adam, and the learning rate was set to [value missing]. .

[0013] For terrain data, the system calculates the slope of each grid cell based on the DEM. slope direction And introduce terrain shading factor ,in The solar azimuth angle is calculated in real time by an astronomical algorithm.

[0014] Based on historical fire point data, the system constructs a spatiotemporal kernel density estimation (ST-KDE) model to calculate the current time for each grid cell. Historical fire point activity ,in and These are the time and space Gaussian kernel functions, and the bandwidth parameter, respectively. , .

[0015] After the above processing, each 100m × 100m grid cell generates an 18-dimensional feature vector in every 10-minute time slice. The components are, in order: NDVI, BAI, Wind speed, wind direction, relative humidity, temperature, precipitation, , , Slope, aspect Historical fire point activity Surface cover type coding (unique thermal vector, occupying 3 dimensions), line voltage level coding (1 dimension, with values ​​1 (500kV), 2 (750kV), 3 (1000kV)).

[0016] In the fusion modeling layer, the system adopts a hybrid architecture combining a Graph Attention Network (GAT) and a spatiotemporal convolutional module. First, all grid cells along the transmission channel are treated as graph nodes, and the node features are the aforementioned 18-dimensional vectors. The edges between nodes are defined by spatial adjacency and transmission topology: if two grid cells are less than 300 meters apart in Euclidean distance, or are located within adjacent spans of the same transmission corridor, then an undirected edge is established. This constructs the graph structure. ,in For a set of nodes, Let it be the set of edges.

[0017] The graph attention layer consists of a two-layer stacked structure. In the first layer, nodes... To his neighbors The attention coefficient is calculated as follows:

[0018] ;

[0019] ;

[0020] ;

[0021] in, For learnable weight matrix, For attention vectors, This represents vector concatenation. This is the ELU activation function.

[0022] The second layer employs a multi-head attention mechanism, setting up four attention heads, each outputting a 32-dimensional vector. After concatenation and linear projection, a 64-dimensional node embedding is obtained. .

[0023] Subsequently, the system introduces a spatiotemporal convolution module to process the time series. It embeds graphs of six consecutive time slices (i.e., 60 minutes) into the sequence. The input is fed into a Causal Dilated Convolution Network. This network consists of three convolutional layers with dilation rates of 1, 2, and 4, a kernel size of 3, 64 channels, and the GELU activation function. The output is the current time step. Spatiotemporal context-enhanced embedding Ultimately, risk prediction is accomplished by a fully connected classification head: ,in Represents grid cells At any moment The probability distribution is based on fire risk levels 0 (very low), 1 (low), 2 (medium), 3 (high), and 4 (very high).

[0024] At the risk output layer, the system selects... The final risk level of the grid cell is determined, and a GeoTIFF format risk level raster map is generated with a spatial resolution of 100 meters and timestamps accurate to the minute. When the risk level of any grid cell reaches level 3 or above for two consecutive time slices, the system automatically sends a level 1 warning signal to the dispatch center, along with the cell's center coordinates, risk probability distribution, dominant feature contribution (calculated through Grad-CAM), and suggested remedial measures (such as drone inspection, line load reduction, manual inspection, etc.).

[0025] In a preferred embodiment of the present invention, the graph attention network weights in the cross-modal feature adaptive fusion prediction module , and classification header parameters , The dataset was obtained through end-to-end training. The training dataset consists of historical data from power transmission channels covering 12 typical fire-prone areas across the country over the past three years, containing over 2 million labeled samples. Each sample is labeled based on the following criteria: if a verified fire occurs within the next 60 minutes in that grid cell, it is marked with the corresponding risk level (mapped according to fire intensity); otherwise, it is marked as level 0. The FocalLoss loss function is used to mitigate class imbalance. ,in The predicted probability corresponding to the true category. , The optimizer uses AdamW with an initial learning rate of... The weight decay coefficient is The batch size is 256, the training rounds are 100, and the early stopping strategy is based on the validation set macro F1-score.

[0026] Furthermore, to enhance the model's robustness under extreme weather events, the system introduces an adversarial example enhancement mechanism during the training phase. Specifically, this mechanism enhances the input feature vector... Add to satisfy disturbance ,in Furthermore, worst-case perturbations are generated using the Projected Gradient Descent (PGD) method, enabling the model to learn invariance to small feature perturbations during training.

[0027] In another preferred embodiment of the present invention, the historical fire behavior pattern mining module employs an improved spatiotemporal point process model. In addition to the standard ST-KDE, the system also incorporates the self-excitation characteristics of fire points modeled using the Hawkes process, with the conditional intensity function being: ,in Based on background strength, For time decay kernel, For spatial decay kernel, parameters Learning is achieved from historical data using maximum likelihood estimation. This intensity function, after normalization, serves as an additional feature. The input is fed into the feature vector, replacing the original ST-KDE output, to more accurately characterize the clustering and spread trends of fire points.

[0028] Furthermore, the video surveillance terminal is deployed on the top of key towers, with a field of view of 120° horizontally and 60° vertically, a lens focal length of 8mm, and equipped with an 850nm infrared LED supplementary light. The LED's center wavelength is 850nm, its radiation intensity is 850mW / sr, and its half-intensity angle is ±25°, ensuring an effective detection distance of no less than 200 meters at night. The video stream, after H.265 encoding, is transmitted to the edge computing node via a 4G / 5G wireless network with a latency of no more than 2 seconds.

[0029] In another preferred embodiment of the present invention, the distributed optical fiber temperature measurement unit adopts the Raman scattering principle, with a laser pulse width of 10 ns, a spatial resolution of 1 meter, a temperature measurement range of -40℃ to +300℃, and a sampling frequency of 0.1Hz. Temperature measurement data is transmitted back via the built-in optical fiber of the OPGW (optical fiber composite overhead ground wire), using the IEC61850-9-2LE communication protocol to ensure seamless integration with the substation automation system.

[0030] Furthermore, the meteorological data access module is equipped with local WRF model microscale nesting capability. When the national WRF forecast update is delayed by more than 30 minutes, the system automatically starts a local 1km resolution nested simulation. The initial field and boundary conditions are provided by the GFS (Global Forecast System) global model, and the physical schemes selected are the YSU planetary boundary layer scheme, the Noah land surface process scheme, and the WSM6 microphysics scheme, with an integration step size of 30 seconds to ensure continuous supply of meteorological elements.

[0031] In another preferred embodiment of the present invention, the risk level output supports dynamic threshold adjustment. The system incorporates a climatological correction factor. Its calculation is based on the sliding Z-score of the regional average FFMC and NDVI over the past 30 days: ,when At that time, the system automatically lowers the judgment thresholds for risk levels 3 and 4 by 10% to cope with the overall risk baseline shifting upward due to the continuous drought and high temperatures.

[0032] Furthermore, the system is deployed at a provincial power grid dispatch center. The hardware platform includes at least four GPU servers (each configured with eight NVIDIA A100 80GB GPUs), ten CPU compute nodes (Intel Xeon Platinum 8380, 40 cores / 80 threads), and a distributed storage cluster (total capacity ≥ 2PB, read / write bandwidth ≥ 10GB / s). The software stack is based on Kubernetes container orchestration, the data pipeline uses the Apache Kafka stream processing framework, and the model inference service is optimized through TensorRT, with end-to-end processing latency controlled within 90 seconds.

[0033] As a key innovation of this invention, the cross-modal feature adaptive fusion mechanism abandons traditional weighted averaging or simple splicing strategies. Instead, it explicitly models spatial dependencies through graph structures and captures dynamic evolution through spatiotemporal convolution, enabling the nonlinear interaction between vegetation dryness, wind-driven conditions, terrain guidance, and historical fire behavior to be resolved within a unified framework. For example, in the windward region of a steep slope, even with a high NDVI, if high wind speed, low humidity, and recent fire records coexist, the model can still enhance the coupling effect of these features through an attention mechanism, outputting a high-risk level, thereby avoiding the misjudgment of "green but dry" vegetation by static overlay methods.

[0034] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0035] This invention achieves high-precision, fine-grained, and real-time prediction of fire risks in power transmission channels by constructing a unified spatiotemporal benchmark for multi-source heterogeneous data, designing a feature engineering process with clear physical meaning, adopting a hybrid modeling paradigm of graph neural networks and spatiotemporal convolution, and integrating dynamic threshold adjustment and adversarial robustness enhancement mechanisms. It effectively overcomes the inherent defects of existing technologies in terms of feature fragmentation, shallow fusion, and time lag, and provides reliable technical support for the active defense system of power grids. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention;

[0037] Figure 2 This is a schematic diagram of the multi-source heterogeneous data access and processing flow in the data acquisition layer of the system of the present invention;

[0038] Figure 3 This is a logical block diagram of the feature engineering layer in the system of the present invention for spatiotemporal alignment and feature vector generation of five types of raw data;

[0039] Figure 4 This is a schematic diagram of the hybrid architecture of graph attention network and spatiotemporal convolution used in the fusion modeling layer of this invention;

[0040] Figure 5 This is a schematic diagram of the 100m × 100m grid cells constructed along the power transmission channel by the system of the present invention and their graphical adjacency relationships;

[0041] Figure 6 This is a schematic diagram illustrating the workflow of generating a fire risk level grid map and triggering graded early warnings for the risk output layer of this invention. Detailed Implementation

[0042] The specific embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit this disclosure.

[0043] In this disclosure, unless otherwise stated, directional terms such as "upper," "lower," "front," "rear," "left," and "right" are used for ease of description based on the drawing orientations of the corresponding figures, while "inner" and "outer" are defined based on the contours of the corresponding components themselves. Terms such as "first" and "second" used in this disclosure are used to distinguish one element from another and do not have sequential or importance implications. Furthermore, when the following description refers to the figures, unless otherwise indicated, the same numbers in different figures represent the same or similar elements.

[0044] This invention provides a feature fusion-based fire risk level prediction system for power transmission channels. Its purpose is to achieve minute-level accurate prediction of fire risk levels within spatial units of hundreds of meters along the power transmission channel through deep fusion and dynamic modeling of multi-source heterogeneous data. The technical solution of this invention will be described in detail, completely, and reproducibly below, with reference to the accompanying drawings and specific engineering implementation details.

[0045] In terms of overall architecture, the system consists of four logical layers: a data acquisition layer, a feature engineering layer, a fusion modeling layer, and a risk output layer, and is deployed on a dedicated hardware platform at the provincial power grid dispatch center. The hardware platform includes at least four GPU servers (each configured with eight NVIDIA A100 80GB GPUs), ten CPU compute nodes (Intel Xeon Platinum 8380, 40 cores / 80 threads), and a distributed storage cluster with a total capacity of no less than 2PB and a read / write bandwidth of no less than 10GB / s. The software stack is built on the Kubernetes container orchestration system, data stream processing uses Apache Kafka as the core message middleware, and the model inference service is optimized with TensorRT to ensure that the end-to-end processing latency from raw data input to risk level output is controlled within 90 seconds.

[0046] The data acquisition layer is responsible for accessing and preprocessing five core data sources: high-resolution remote sensing imagery, numerical weather prediction data, online monitoring data of power transmission lines, digital elevation models and land use data, and historical fire point records. High-resolution remote sensing imagery comes from multispectral imagers mounted on low-Earth orbit satellites or relay UAV platforms. Its spatial resolution ranges from 0.5 to 2 meters, with a revisit time of no more than 6 hours, and its spectral bands cover visible light (450–700 nm), near-infrared (700–1000 nm), and shortwave infrared (1000–2500 nm). Numerical weather prediction data comes from the output of the WRF (Weather Research and Forecasting) model released by the National Meteorological Center. Its spatial grid accuracy is 1 km × 1 km, and its time step is 1 hour. It includes wind speed (m / s), wind direction (°), relative humidity (%), temperature (°C), precipitation (mm / h), and flammability dryness index (FFMC). The online monitoring data for transmission lines originates from distributed fiber optic temperature measurement units, current transformers, and video monitoring terminals deployed on the towers. The distributed fiber optic temperature measurement units utilize Raman scattering principles, with a laser pulse width of 10 ns, a spatial resolution of 1 meter, a temperature measurement range of -40℃ to +300℃, and a sampling frequency of 0.1Hz. Temperature data is transmitted back via the built-in fiber optic cable of the OPGW (Optical Fiber Composite Overhead Ground Wire), and the communication protocol conforms to the IEC61850-9-2LE standard. The video monitoring terminals are deployed on the tops of key towers, with a field of view of 120° horizontally and 60° vertically, an 8mm lens focal length, and equipped with 850nm infrared LED supplementary lights. The LEDs have a center wavelength of 850nm, a radiation intensity of 850mW / sr, and a half-intensity angle of ±25°, ensuring an effective detection distance of at least 200 meters at night. The video stream is encoded using H.265 and transmitted to edge computing nodes via a 4G / 5G wireless network, with an end-to-end transmission delay of no more than 2 seconds. The digital elevation model (DEM) and land use data are sourced from the 1:50,000 scale geographic database published by the National Geomatics Center of China, with a horizontal positioning error of less than 5 meters. Historical fire point records are sourced from the fire monitoring platform of the National Forestry and Grassland Administration, including verified fire point coordinates, ignition time, burned area, and fire intensity level over the past five years, with timestamp accuracy at the minute level and a spatial positioning error of less than 30 meters.

[0047] Before entering the feature engineering layer, all raw data undergoes a unified spatiotemporal normalization operation by the spatiotemporal reference alignment module. Spatially, all data is projected onto the WGS84 geographic coordinate system and resampled to a unified 100m × 100m regular grid cell. Each grid cell corresponds to a basic risk assessment unit along the transmission channel, such as... Figure 5As shown. In the time dimension, all data streams are based on UTC timestamps and are aligned to a time slice of 10 minutes through linear interpolation or nearest neighbor padding to form a continuous time series data stream.

[0048] In the feature engineering layer, the five types of raw data are processed by five dedicated sub-modules to generate feature components with clear physical meaning, which are then concatenated into an 18-dimensional feature vector.

[0049] The remote sensing feature extraction submodule is responsible for processing high-resolution remote sensing image data. For each 100m × 100m grid cell, the system calculates the Normalized Difference Vegetation Index (NDVI), which is defined as follows: ,in and These are the surface reflectance values ​​for the near-infrared and red light bands, respectively. The burn area index (BAI) is also calculated using the following formula: Surface temperature ( The formula is obtained through inversion using a single-window algorithm. ,in For thermal infrared radiance, a constant , .

[0050] The meteorological feature correction submodule processes numerical weather prediction data. The system directly extracts the WRF output value corresponding to each grid cell and further calculates the corrected combustible dryness index. ,in The relative humidity gradient is calculated by dividing the relative humidity difference between adjacent grid cells by the Euclidean distance, and is used to reflect the abrupt changes in local microclimate. Furthermore, when the national WRF forecast update is delayed by more than 30 minutes, the system automatically initiates a local 1km resolution nested simulation. The initial field and boundary conditions are provided by the GFS (Global Forecast System) global model, and the physical schemes used are the YSU planetary boundary layer scheme, the Noah land surface process scheme, and the WSM6 microphysics scheme, with an integration step size of 30 seconds to ensure a continuous supply of meteorological data.

[0051] The line condition mapping submodule processes online monitoring data for transmission lines. Fiber optic temperature measurement sequence. The tower locations are mapped to corresponding grid cells, and the average conductor temperature within each cell is calculated. ,in This represents the number of temperature measurement points falling into the grid. Smoke confidence scores are extracted from video surveillance data using a pre-trained convolutional neural network (ResNet-18 architecture). After pre-training on the ImageNet large-scale image dataset, the network was fine-tuned using a power scene-specific dataset containing 100,000 labeled smoke / non-smoke images. The loss function was cross-entropy, the optimizer was Adam, and the learning rate was set to [value missing]. .

[0052] The terrain factor calculation submodule processes DEM and land use data. The system calculates the slope of each grid cell based on the DEM. slope direction And introduce terrain shading factor ,in The solar azimuth angle is calculated in real time by an astronomical algorithm based on the current UTC time and geographical location. Land use types are encoded as 3-dimensional unique heat vectors, corresponding to major combustible material categories such as forest, shrub, and grassland.

[0053] The historical fire behavior mining submodule processes historical fire point record data. In a preferred embodiment of the invention, this module first constructs a standard spatiotemporal kernel density estimation (ST-KDE) model to calculate the historical fire point record data for each grid cell at the current time. Historical fire point activity ,in and All are Gaussian kernel functions, with bandwidth parameters. , As another preferred implementation, this module further incorporates the self-excitation characteristics of fire points modeled by the Hawkes process, with its conditional intensity function being: Among them, the basic background intensity The time decay kernel is a constant. Spatial decay kernel ,parameter Learning is achieved from historical data using maximum likelihood estimation. This intensity function, after normalization, serves as an additional feature. Input, replacing the original ST-KDE output.

[0054] After processing by the above five sub-modules, each 100m × 100m grid cell generates an 18-dimensional feature vector in every 10-minute time slice. The components are, in order: NDVI, BAI, Wind speed, wind direction, relative humidity, temperature, precipitation, , , Slope, aspect Historical fire point activity Surface cover type coding (unique thermal vector, occupying 3 dimensions), line voltage level coding (1 dimension, with values ​​1 (500kV), 2 (750kV), 3 (1000kV)).

[0055] The fusion modeling layer receives the aforementioned feature vectors and constructs a hybrid architecture combining a graph attention network (GAT) and a spatiotemporal convolution module for deep feature fusion and dynamic prediction. First, all grid cells along the transmission channel are considered as graph nodes, and the node features are 18-dimensional vectors. The edges between nodes are defined by spatial adjacency and transmission topology: if two grid cells are less than 300 meters apart in Euclidean distance, or are located within adjacent spans of the same transmission corridor, then an undirected edge is established, such as... Figure 5 The spatial adjacent edges and transmission topology edges are shown in the diagram. This leads to the construction of a graph structure. .

[0056] The graph attention network module consists of a two-layer stacked structure. In the first layer, nodes... To his neighbors The attention coefficient is calculated as follows:

[0057] ;

[0058] ;

[0059] ;

[0060] in For learnable weight matrix, For attention vectors, This represents vector concatenation. This is the ELU activation function.

[0061] The second layer employs a multi-head attention mechanism, setting up four attention heads. Each head independently calculates attention weights and outputs a 32-dimensional vector. The outputs of the four heads are then concatenated into a 128-dimensional vector, and then processed through a... The linear projection layer yields the final 64-dimensional node embedding. .

[0062] Subsequently, the spatiotemporal convolution module handles the dynamic evolution along the time dimension. The system embeds a sequence of graphs from six consecutive time slices (i.e., 60 minutes). The input is fed into a Causal Dilated Convolution Network. This network consists of three one-dimensional convolutional layers with dilation rates of 1, 2, and 4, a kernel size of 3, and 64 channels in each layer. The GELU activation function is used. Causality is ensured by applying a mask to the convolutional kernels, so that the output at the current time step depends only on the past and current inputs, without revealing future information. The output of this module is the current time step. Spatiotemporal context-enhanced embedding .

[0063] Finally, the fully connected classification head receives... It also outputs the probability distribution of fire risk levels: ,in Represents grid cells At any moment The probability of being in a fire risk level of 0 (very low), 1 (low), 2 (medium), 3 (high), or 4 (very high).

[0064] As a key implementation detail of this invention, the entire fusion model is obtained through end-to-end training. The training dataset consists of historical data from power transmission channels covering 12 typical fire-prone areas across the country over the past three years, containing more than 2 million labeled samples. Each sample is labeled based on the following criteria: if a verified fire occurs within the grid cell within the next 60 minutes, it is mapped to a risk level of 2, 3, or 4 according to the fire intensity (determined by a combination of burned area and spread rate); otherwise, it is labeled as level 0. Due to the extreme imbalance between positive and negative samples (high-risk samples account for less than 0.8%), the FocalLoss loss function is used. ,in The hyperparameter represents the predicted probability corresponding to the true class. , The optimizer uses AdamW with an initial learning rate of... The weight decay coefficient is The batch size is 256, the training rounds are 100, and an early stopping strategy is adopted, which terminates training when the validation set macro F1-score does not improve within 10 consecutive rounds.

[0065] To further enhance the model's robustness under extreme weather events, the system introduces an adversarial example enhancement mechanism during the training phase. Specifically, this mechanism enhances the input feature vector... Add to satisfy disturbance ,in Furthermore, worst-case perturbations are generated iteratively using the Projected Gradient Descent (PGD) method, enabling the model to learn invariance to small feature perturbations during training, thereby enhancing its adaptability to sensor noise or data drift in actual operation.

[0066] At the risk output layer, the system selects... The final risk level for this grid cell is used to generate a risk level raster map in GeoTIFF format with a spatial resolution of 100 meters and timestamps accurate to the minute. In a preferred embodiment of the invention, the system incorporates a dynamic threshold adjustment module, calculated based on the sliding Z-score of the regional average FFMC and NDVI over the past 30 days. ,in The mean and standard deviation of the regional FFMC over the past 90 days. Similarly. When At that time, the system automatically lowers the judgment thresholds for risk levels 3 and 4 by 10% to cope with the overall risk baseline shifting upward due to the continuous drought and high temperatures.

[0067] When the risk level of any grid cell reaches level 3 or above for two consecutive time slices (i.e., 20 minutes), the system automatically sends a level 1 warning signal to the dispatch center and pushes it to relevant operation and maintenance terminals via API interface. The warning information package includes the center coordinates of the cell and the risk probability distribution. The contribution of dominant features and recommended remedial measures are as follows: The contribution of dominant features is calculated using Gradient Weighted Class Activation Mapping (Grad-CAM) technology. Specifically, the gradients of the classification head output with respect to each component of the input feature vector are globally averaged and pooled, then weighted and summed with the original feature map to quantify the contribution weight of each feature to the final prediction. Recommended remedial measures are automatically generated based on the risk level and dominant features. For example: If... If the contribution is the highest, it is recommended to "immediately dispatch drones for close-range reconnaissance"; if If the combined effect of wind speed and other factors is high, it is recommended to "implement load reduction operation on the relevant sections of the line"; if the activity of historical fire points and the gradient are dominant, it is recommended to "deploy additional manual inspection teams".

[0068] This invention has been described through embodiments. Those skilled in the art will understand that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of this invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, this invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of this invention.

Claims

1. A feature fusion-based power transmission channel fire risk level prediction system, characterized in that, It includes a data acquisition layer, a feature engineering layer, a fusion modeling layer, and a risk output layer; The data acquisition layer is used to access high-resolution remote sensing image data, numerical weather forecast data, online monitoring data of power transmission lines, digital elevation models and land use data, and historical fire point records. The feature engineering layer includes a spatiotemporal reference alignment module, which projects all raw data to the WGS84 geographic coordinate system and resamples it into 100m×100m regular grid cells. At the same time, it aligns the data to a time slice every 10 minutes based on the UTC timestamp and generates an 18-dimensional feature vector corresponding to each grid cell. The spatiotemporal benchmark alignment module includes a remote sensing feature extraction submodule, a meteorological feature correction submodule, a line condition mapping submodule, a terrain factor calculation submodule, and a historical fire behavior mining submodule. The fusion modeling layer includes a graph attention network module and a spatiotemporal convolution module. The graph attention network (GAT) module constructs a graph structure with each grid cell as a node and edges defined by spatial adjacency and transmission topology. It generates 64-dimensional node embeddings through a two-layer stacked graph attention mechanism. The spatiotemporal convolution module performs causal dilation convolution processing on the node embedding sequence of six consecutive time slices and outputs spatiotemporal context-enhanced embeddings. The risk output layer includes a fully connected classification head, which is used to enhance the probability distribution of the embedded output fire risk level based on the spatiotemporal context.

2. The feature fusion based power transmission corridor fire risk rating prediction system of claim 1, wherein, The 18-dimensional feature vector includes: Normalized Difference Vegetation Index (NDVI), Burning Area Index (BAI), Surface Temperature (Tsurf), Wind Speed, Wind Direction, Relative Humidity, Air Temperature, Precipitation, Corrected Combustion Dryness Index (FFMCadj), and Average Conductor Temperature. Smoke confidence score Slope, aspect, and topographic shielding factor Historical fire point activity Surface cover type coding (3-dimensional unique thermal vector) and line voltage level coding (1-dimensional). 3.The feature fusion based power transmission corridor fire risk level prediction system of claim 1, wherein, An undirected edge is established if any of the following conditions are met: the two grid cells are less than 300 meters apart in Euclidean distance, or they are located within adjacent spans of the same power transmission corridor.

4. The feature fusion based power transmission right-of-way fire risk rating prediction system of claim 1, wherein, The first layer of the graph attention network (GAT) module adopts a single-head attention mechanism, and the second layer adopts a 4-head multi-head attention mechanism. The outputs of each head are spliced ​​together and linearly projected to obtain a 64-dimensional node embedding.

5. The feature fusion based power transmission right-of-way fire risk rating prediction system of claim 1, wherein, The spatiotemporal convolution module contains three causal dilated convolutional layers with dilation rates of 1, 2, and 4, respectively. The kernel size is 3, the number of channels is 64, and the activation function is GELU. 6.The feature fusion based power transmission right-of-way fire risk level prediction system of claim 1, wherein, The historical fire behavior mining submodule uses a spatiotemporal kernel density estimation model to calculate the activity of historical fire points. Its bandwidth parameter is time bandwidth. Hours, spatial bandwidth Meters; or the conditional intensity function can be calculated using the Hawkes process model. Among them, time decay kernel Spatial decay kernel ,parameter Learn from historical data through maximum likelihood estimation.

7. The power transmission channel fire risk level prediction system based on feature fusion according to claim 1, characterized in that, The online monitoring data for the transmission lines includes temperature data collected by distributed fiber optic temperature measurement units and video streams collected by video monitoring terminals. The distributed fiber optic temperature measurement units adopt the Raman scattering principle, with a spatial resolution of 1 meter, a temperature measurement range of -40℃ to +300℃, and a sampling frequency of 0.1Hz. The data is transmitted back via the OPGW's built-in fiber optic cable using the IEC 61850-9-2LE protocol. The video monitoring terminals are deployed on the top of key towers, equipped with 850nm infrared LED supplementary lights, and have a field of view of 120° horizontally and 60° vertically. The video stream is encoded with H.265 and transmitted to the edge computing node via a 4G / 5G network, with an end-to-end latency of no more than 2 seconds. 8.The feature fusion based power transmission right-of-way fire risk level prediction system of claim 1, wherein, The meteorological feature correction submodule is equipped with local WRF model microscale nesting capability. When the national WRF forecast update is delayed by more than 30 minutes, a 1km resolution nested simulation is automatically started. The initial field and boundary conditions are provided by the GFS global model. The physical schemes include the YSU planetary boundary layer scheme, the Noah land surface process scheme, and the WSM6 microphysics scheme, with an integration step size of 30 seconds. 9.The feature fusion based power transmission right-of-way fire risk level prediction system of claim 1, wherein, The risk output layer further comprises a dynamic threshold adjustment module for adjusting the risk level determination threshold dynamically according to the climate state correction factor The determination threshold of the risk levels 3 and 4 is lowered by 10% when The determination threshold of the risk levels 3 and 4 is lowered by 10% when 10. The feature fusion based power transmission right-of-way fire risk rating prediction system of claim 1, wherein, When the risk level of any grid cell reaches level 3 or above for two consecutive time slices, the system automatically sends a level 1 early warning signal to the dispatch center, along with the cell's center coordinates, risk probability distribution, dominant feature contribution, and suggested handling measures. The dominant feature contribution is calculated using Grad-CAM technology, and the suggested handling measures are automatically generated based on the dominant features, including UAV inspection, line load reduction, or manual inspection.

Images