Multi-factor coupling power transmission line forest fire risk early warning method
By employing a multi-factor coupled early warning method, combining Himawari satellite and ground meteorological data, and utilizing the GBDT-GWR framework to construct an early warning model that integrates global and local data, the problem of differences in the microenvironment and heterogeneity of fire risk factors at the pole level in power transmission line wildfire early warning was solved, achieving high-precision pole-level risk assessment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA THREE GORGES UNIV
- Filing Date
- 2026-01-29
- Publication Date
- 2026-05-19
AI Technical Summary
Existing power transmission line wildfire early warning technologies cannot accurately reflect the differences in the microenvironment at the tower level, ignore the spatial heterogeneity of fire risk factors, and lack a multi-level early warning framework, resulting in insufficient early warning accuracy.
A multi-factor coupled early warning method is adopted, combining Himawari satellite data and ground meteorological data. Through gradient boosting decision tree regression analysis and geographically weighted regression, an early warning model combining global and local aspects is constructed. The GBDT-GWR coupling framework is used for fire risk factor classification and risk assessment.
It achieves risk stratification at the 1 km × 1 km grid level, improving accuracy by 23.4%, and can dynamically correct tower-level risks in transmission line corridors, making it suitable for wildfire early warning in complex terrain.
Smart Images
Figure CN122067360A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system technology, and in particular to the operation and disaster prevention and mitigation technology of power systems. Specifically, it relates to a "global early warning-local correction" method for early warning of wildfire risks in transmission line corridors based on multi-factor coupling. Background Technology
[0002] With global warming and the increase in extreme weather events, frequent wildfires have become a significant threat to the safe operation of power transmission lines. Wildfires occurring within transmission line corridors can not only cause air gap breakdowns and protective tripping, but may also lead to large-scale power outages, endangering the stable operation of the power system.
[0003] Existing methods for wildfire risk early warning mainly include: 1. Global monitoring based on meteorology and remote sensing, using satellite data such as MODIS and VIIRS to identify fire points over a wide area, and constructing forest fire weather risk level models based on the time frequency of fire point occurrences. However, its spatial resolution is usually at the kilometer level, which is insufficient to meet the refined early warning needs at the pole and tower level. 2. Fire risk modeling based on statistics or machine learning: using methods such as logistic regression, random forest, and Boosting to model fire risk factors, assigning summation weights to different factors, and quantifying the fire weather risk level. However, this method has shortcomings when it does not consider the spatial heterogeneity of fire risk factors and the localized characteristics of transmission line poles and towers. 3. Specialized research on transmission lines: some methods introduce methods such as UAV inspections and local vegetation burning tests to classify risk levels, but the coverage is limited, making it difficult to form a coordinated early warning system at the regional and pole / tower levels.
[0004] More specifically, in existing technologies, for example, the article "Research on Forest Fire Prediction and Monitoring Based on Multi-Factor Synergistic Effect of MODIS Data" describes a method for forest fire prediction and monitoring based on MODIS remote sensing data and integrating multiple meteorological factors. However, this technology suffers from technical limitations due to the fact that the spatial resolution of MODIS data is usually at the kilometer level, making it difficult to characterize the complex terrain, vegetation, and local differences in micro-meteorological conditions along transmission lines. As a result, when applied to forest fire risk early warning, it is difficult to achieve precise identification of fire risks at the tower level, and local high-risk areas are easily smoothed or underestimated, thus affecting the accuracy of the early warning.
[0005] In summary, current early warning systems for wildfires in power transmission corridors still have the following problems: insufficient resolution of global early warning models, which cannot accurately reflect the differences in microenvironments at the tower level; significant spatial heterogeneity of fire risk factors, with traditional global models ignoring local differences and affecting early warning accuracy; and a lack of a multi-level early warning framework that combines large-scale remote sensing monitoring with on-site characteristics of power transmission corridors.
[0006] Therefore, how to accurately issue early warnings of wildfires along power transmission line corridors remains an urgent problem to be solved. Summary of the Invention
[0007] The purpose of this invention is to address the following problems in existing early warning technologies for power transmission corridors: insufficient resolution of the global early warning model, failing to accurately reflect the differences in the microenvironment at the tower level; significant spatial heterogeneity of fire risk factors, with traditional global models ignoring local differences and affecting early warning accuracy; and the lack of a multi-level early warning framework that combines large-scale remote sensing monitoring with on-site characteristics of power transmission corridors. These technical problems make existing early warning technologies for power transmission corridors increasingly unable to meet the needs of early warning for power transmission corridors, hence this invention.
[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A multi-factor coupled method for early warning of wildfire risks along power transmission lines, comprising the following steps: Step 1: Collect data on several historical wildfire hotspots in the target area, and obtain meteorological data such as monthly precipitation, monthly average temperature, soil moisture and wind speed in the area. At the same time, obtain surface reflectance data affected by hyperspectral remote sensing to calculate vegetation ecological factors and perform land use classification, and construct a standardized multi-factor fire risk factor dataset for wildfire risk analysis. Step 2: Based on the multi-factor fire risk factor dataset obtained in Step 1, the fire risk factors are classified according to their physical meaning and risk change characteristics. On this basis, a sample dataset for global wildfire risk modeling is constructed, and then the model is trained and the hyperparameters are tuned to finally obtain the regional scale global fire risk level distribution results. Step 3: Based on the regional-scale global fire risk level distribution results obtained in Step 2, obtain the fire risk distribution map of transmission corridor towers; Step 4: Based on the results of the distribution of forest fire risk at the pole level obtained in Step 3, set the threshold for classifying fire risk levels and classify the forest fire risk level of each pole along the transmission line.
[0009] In step 1, data on several historical wildfire hotspots in the target study area over a period of time were collected from Himawari geostationary satellite observations. Simultaneously, meteorological data such as monthly precipitation, monthly average temperature, soil moisture, and wind speed were acquired at a designated scientific data center. Meanwhile, surface reflectance data influenced by hyperspectral remote sensing were obtained from the Natural Resources Remote Sensing Cloud Service Platform to calculate NDVI vegetation ecological factors and perform land use classification. In ArcGIS, the projection raster tool was used to convert the projection coordinate system of all factors to WGS_1984_UTM_Zone_50N. The resampling tool was used to uniformly set the cell size of all raster data to 1 km², constructing a standardized multi-factor fire risk factor dataset for wildfire risk analysis.
[0010] In step 1, vegetation ecological factors include the Normalized Difference Vegetation Index (NDVI), vegetation cover (FVC), digital elevation model (DEM), slope, and aspect. Characteristic factors of power transmission corridors include minimum clearance distance between transmission line towers, key governance areas in an inverted triangle shape, distance from historical fire points, and distribution of cemeteries.
[0011] In step 2, the process for conducting a global wildfire risk early warning analysis is as follows: Step 2-1) Using several fire points in the target month as targets, overlay multiple fire risk factor data onto each fire point and classify the fire risk factors. Step 2-2) Using the combination of warning index levels of each fire risk factor as the independent variable, the number of all fire points under the combination of levels is counted as the dependent variable. Using Python, gradient boosting decision tree regression analysis based on feature selection and Bayesian parameter tuning is used to quantitatively evaluate the importance weight of each fire risk factor. Steps 2-3) Obtain the fire risk warning value. The fire risk warning value is a weighted linear representation of the characteristic contribution degree of the fire risk factor and the fire risk factor level. The fire risk warning value is calculated by weighting the importance of each feature obtained for several fire points in the target area. The fire risk warning value of the entire area is obtained by using the Kriging interpolation method based on this.
[0012] In step 2-2), when performing gradient boosting decision tree regression analysis to quantify the importance weights of each fire risk factor, the following steps are adopted: 1) Initialization, using the target variables in the training set. The average number of fire points is used as the initial prediction value. In the formula: Represents the total number of samples. For the first The number of fire points per sample, initializing the prediction function. As the baseline prediction function for the first iteration of the gradient boosting decision tree model, it is used for subsequent residual calculation and model update;
[0013] 2) Iterative training process, in the first... In the next iteration, the predicted value is calculated. Compared with the true value The residual (negative gradient), where when m=1, the This is the initial prediction function obtained in step 1. :
[0014] In the formula, It is the objective function. The model is for the first The prediction function for each sample. It is the first After the nth iteration, the model... For each sample, the prediction result, for the mean squared error, the residual is equivalent to the difference between the current error value and the true value:
[0015] 3) Utilizing residuals As a pseudo-target value, train a new learner. Fit the current residual distribution; combine the current learner's predictions to update the model's predictions:
[0016] In the formula: It is the first The warning function at the next iteration It is the learning rate, which controls the step size of each update to avoid overfitting. It is in the The learner, trained using the residual as a pseudo-target value in each iteration, continuously optimizes the predicted values through multiple iterations. Until the residual converges and the region of change of the objective function stabilizes; 4) Use the Bayesian optimizer to perform hyperparameter tuning (number of decision trees, learning rate, maximum depth of decision trees) to obtain the optimal operating parameters of the GBDT classifier; 5) To quantitatively assess the relative importance of each fire risk factor to the early warning model, based on the feature importance calculation mechanism built into GBDT, a feature contribution evaluation system is constructed by statistically analyzing the average information gain (measured by the reduction in mean squared error) brought by each feature when splitting all regression tree nodes. Equation (5) is used to quantify the importance weight of each feature:
[0017] in Indicates the first One characteristic, For the total number of decision trees, For the first Used in trees The set for which nodes are split. This represents the decrease in mean square error caused by the split point S.
[0018] In step 3, the following steps are performed: Step 3-1) Perform preliminary GWR modeling; within the target area, establish a preliminary geographically weighted regression GWR model for fire risk factors and fire risk warning values, embed the regression coefficients into the spatial coordinates of the sample points to obtain the local regression coefficients and local fitting values of each sample point, providing basic data for subsequent spatial feature analysis; Step 3-2) Perform spatial autocorrelation analysis; Step 3-3) Perform multicollinearity diagnosis; After completing the spatial autocorrelation analysis in Step 3-2 and confirming that the dependent variable fire risk warning value has significant spatial heterogeneity, perform multicollinearity test on the independent variables of the GWR model to ensure the stability and explanatory power of the model. Step 3-4) Obtain the local regression coefficients of the GWR fire risk factor; specifically, after completing the multicollinearity screening in step 3-3, use the optimized combination of independent variables to perform the final fitting of GWR.
[0019] In step 3-1), when performing preliminary GWR modeling, the independent variables not only include the original fire risk factors such as meteorology, remote sensing, and topography, but also introduce on-site factors of transmission lines, including the minimum clearance distance of transmission line towers, the inverted triangle key treatment area, and the shortest distance from each tower to the fire point. The dependent variable is the fire risk warning value, which is derived from the global fire risk warning map generated by the GBDT method in step 2, and the local risk value of each fire point is obtained through target area masking. By corresponding the fire risk factors with the fire risk value according to spatial coordinates, the GWR model can quantify the local contribution of each factor to the fire risk in different geographical locations, thus providing a basis for subsequent spatial autocorrelation analysis, multicollinearity diagnosis, and final GWR model optimization. Geographically Weighted Regression (GWR) is a regression analysis method that can capture spatial heterogeneity. It allows the regression coefficients to change spatially, considers the local effects of spatial objects, and embeds the geographical location of the sample data into the regression coefficients in the form of distance weights. The expression of the GWR model is shown in equation (6):
[0020] In the formula: It is the first Spatial coordinates of each sample point Located in spatial coordinates The local intercept term at the location is used to characterize the baseline fire risk level at that location without considering the influence of various fire risk factors; Indicates the first The sample point of the th sample point Local regression coefficients of the independent variables, These are independent and identically distributed error terms; the spatial coordinates It is used to characterize the relative positional relationship of sample points in geospace and serves as a unified spatial benchmark for subsequent spatial autocorrelation analysis, spatial weight calculation, and geographic weighted regression parameter estimation.
[0021] In step 3-2), the following formula is used for judgment when performing spatial autocorrelation analysis:
[0022] The total number of observation units, , The first The and the first Fire risk index for each unit, This represents the average of the fire risk index. The spatial weight matrix represents the unit. and unit The spatial adjacency relationship between them, wherein the spatial weight matrix Based on the sample points being in the same spatial coordinate system The spatial autocorrelation of the fire risk index was constructed using distance relationships. Further statistical tests were conducted to verify this. The Moran's index was 0.881, significantly close to 1, indicating a strong positive spatial correlation. High and low values clustered spatially. The Z-value, a standardized value of the observed Moran's index, was 14.89, far exceeding 1.96 (corresponding to a 95% confidence level). Furthermore, the p-value's significance probability was close to 0 and far less than 0.05, supporting the rejection of the null hypothesis. This indicates that the spatial clustering is highly statistically significant. The spatial distribution pattern of the fire risk index is suitable for GWR modeling to explore the spatial heterogeneity of different fire risk factors.
[0023] In step 3-3, the variance inflation factor (VIF) diagnostic method is used to diagnose multicollinearity. The expression for the variance inflation factor is:
[0024] In the formula: as independent variable The multiple correlation coefficients of linear regression analysis performed on the remaining (n-1) independent variables. When When there is no multicollinearity among the variables; when There is strong multicollinearity among the variables. In this invention, the VIF values of all candidate fire risk factors are calculated for preliminary diagnosis. For factors with VIF values significantly higher than the threshold, necessary variable elimination will be performed according to the diagnosis results to ensure that the combination of independent variables finally included in the GWR model meets the requirement of low collinearity, thereby improving the reliability and explanatory power of the model. Through screening, the finally selected independent variables and their VIF values are as follows: fire point distance (1.49), " / \" key area (1.19), NDVI (1.55), slope (3.25), slope aspect (1.27), population density (1.29), land cover classification (1.28), minimum clearance distance (1.09), and wind speed (1.52). All of these are significantly lower than the collinearity threshold, indicating that the correlation between variables is low, which can ensure the fitting stability and prediction accuracy of the GWR model. In steps 3-4, a Gaussian spatial weighting function is introduced to reveal the contribution and mechanism of each fire risk factor at different geographical locations; the expression of the Gaussian spatial weighting function is shown in Equation 9:
[0025] In the formula: Indicates the first A pair of objects The weight of influence For sample points With sample points In spatial coordinates and The Euclidean distance is calculated below. The bandwidth parameter determines the search range within the domain.
[0026] It also includes steps 3-5), optimizing the GWR bandwidth; using spatial statistics tools in ArcGIS to implement GWR regression, considering that the distribution of transmission line towers is not completely uniform, in the GWR regression process, the adaptive neighborhood type is selected and a fixed number of neighborhoods is used as the bandwidth range to ensure that each sample point contains sufficient neighborhood information in the local regression, and the Akaike Information Criterion (AIC) is selected to automatically search for the optimal bandwidth value to achieve local optimal fitting of the model.
[0027] Compared with the prior art, the present invention has the following technical effects: 1) The multi-factor coupled GBDT global model based on the Himawari satellite proposed in this invention (R²=0.626, RMSE=0.178) initially achieves risk stratification at the 1 km×1 km grid level; after local GWR correction, the model fit goodness improved to R²=0.791 (p<0.01), and the accuracy improved by 23.4%; 2) This invention provides a new approach to forest fire early warning research, taking a total of 412 towers of the two Xiaoshi lines as the research point. Compared with traditional forest fire early warning models, this study deeply integrates meteorological data, remote sensing data, field parameters of the transmission corridor, and GIS spatial analysis technology through the GBDT-GWR coupling framework. With historical fire points as a reference, it realizes dynamic correction of tower-level risk of transmission line corridor from two-stage wide-area monitoring to local response, and solves the problem that the global weight of fire risk factors cannot be applied to complex terrain of transmission corridors. 3) This invention can be extended to areas prone to wildfires (such as the Yunnan-Guizhou Plateau and the Jiangnan Hills). Through transfer learning, it can adapt to different geographical limitations and improvement directions. In addition, this invention relies on static historical fire point data and does not integrate real-time meteorological data (such as sudden changes in wind speed and drought index) to consider risk warnings and fire point spread on a short time scale. Attached Figure Description
[0028] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a visualization of the fire hazard factor raster data in an embodiment of the present invention; Figure 3 This is a map showing the distribution of fire points on the Himawari satellite in an embodiment of the present invention. Figure 4 This is a ranking diagram of the contribution of fire hazard factors in this invention; Figure 5 This is a global fire hazard level map of the power transmission corridor in an embodiment of the present invention; Figure 6 This is a transmission line early warning diagram after Kriging interpolation in an embodiment of the present invention; Figure 7 This is a flowchart of the refined correction process for power transmission corridor sections in this invention; Figure 8 This is a modified fire hazard rating diagram of the power transmission corridor in an embodiment of the present invention; Figure 9 This is a diagram showing the specific tower section levels before and after modification in an embodiment of the present invention; Figures 10 to 17 This is a distribution diagram of the regression coefficients of fire risk factors in an embodiment of the present invention. Detailed Implementation
[0029] like Figure 1 As shown, a multi-factor coupled method for early warning of wildfire risks along power transmission lines includes the following steps: Step 1: Collect data on several historical wildfire hotspots in the target area, and obtain meteorological data such as monthly precipitation, monthly average temperature, soil moisture and wind speed in the area. At the same time, obtain surface reflectance data affected by hyperspectral remote sensing to calculate vegetation ecological factors and perform land use classification, and construct a standardized multi-factor fire risk factor dataset for wildfire risk analysis. Step 2: Based on the multi-factor fire risk factor dataset obtained in Step 1, the fire risk factors are classified according to their physical meaning and risk change characteristics. On this basis, a sample dataset for global wildfire risk modeling is constructed, and then the model is trained and the hyperparameters are tuned to finally obtain the regional scale global fire risk level distribution results. Step 3: Based on the regional-scale global fire risk level distribution results obtained in Step 2, obtain the fire risk distribution map of transmission corridor towers; Step 4: Based on the results of the distribution of forest fire risk at the pole level obtained in Step 3, set the threshold for classifying fire risk levels and classify the forest fire risk level of each pole along the transmission line.
[0030] In step 1, data on several historical wildfire hotspots in the target study area over a period of time were collected from Himawari geostationary satellite observations. Simultaneously, meteorological data such as monthly precipitation, monthly average temperature, soil moisture, and wind speed were acquired at a designated scientific data center. Meanwhile, surface reflectance data influenced by hyperspectral remote sensing were obtained from the Natural Resources Remote Sensing Cloud Service Platform to calculate NDVI vegetation ecological factors and perform land use classification. In ArcGIS, the projection raster tool was used to convert the projection coordinate system of all factors to WGS_1984_UTM_Zone_50N. The resampling tool was used to uniformly set the cell size of all raster data to 1 km², constructing a standardized multi-factor fire risk factor dataset for wildfire risk analysis.
[0031] In step 1, the fire risk influencing factors include meteorological factors, vegetation ecological factors, and transmission corridor characteristic factors, wherein: the meteorological factors include monthly total precipitation, monthly average temperature, monthly average wind speed, and monthly average soil moisture; the vegetation ecological factors include normalized difference vegetation index (NDVI), vegetation cover (FVC), enhanced vegetation index (EVI), normalized difference water index (NDWI), bare soil index (BSI), building index (IBI), as well as digital elevation model (DEM), slope, and aspect; the transmission corridor characteristic factors include minimum clearance distance between transmission line towers, inverted triangle key treatment area, distance from historical fire points, distribution of cemeteries, and the proportion of land use types within the preset buffer zone around the towers and the distance from the towers to the nearest roads and residential areas.
[0032] In step 2: a global wildfire risk early warning analysis is performed. Based on the multi-factor fire risk factor dataset obtained in step 1, the fire risk factors are classified into 1–5 according to their physical meaning and risk change characteristics. The study first determined the fire risk level. Based on this, it statistically analyzed the historical number of fire points under different combinations of fire risk factor levels, using each combination of fire risk factor levels as the independent variable and the number of fire points as the dependent variable to construct a sample dataset for global wildfire risk modeling. Fire risk factors were screened using recursive feature elimination and cross-validation methods, selecting a set of fire risk factors that significantly influence the number of fire points. A gradient boosting decision tree regression model was then used to train the sample data, and the model's hyperparameters were adaptively tuned using a Bayesian parameter optimization method. After model training, based on the feature importance calculation mechanism built into the gradient boosting decision tree model, a fire risk factor feature contribution evaluation system was constructed by statistically analyzing the average information gain brought by each fire risk factor during the splitting of all regression tree nodes. This quantified the feature importance weights of each fire risk factor, thus obtaining the calculation formula for the fire risk warning value. The calculated fire risk index was used to classify the fire risk level of the study area, and Kriging interpolation was used to spatially interpolate the fire risk warning value, generating a regional-scale global fire risk level distribution result. In step 2, the process for conducting a global wildfire risk early warning analysis is as follows: Step 2-1) Using several fire points in the target month as targets, overlay multiple fire risk factor data onto each fire point and classify the fire risk factors. Step 2-2) Using the combination of warning index levels of each fire risk factor as the independent variable, the number of all fire points under the combination of levels is counted as the dependent variable. Using Python, gradient boosting decision tree regression analysis based on feature selection and Bayesian parameter tuning is used to quantitatively evaluate the importance weight of each fire risk factor. Steps 2-3) Obtain the fire risk warning value. The fire risk warning value is a weighted linear representation of the characteristic contribution degree of the fire risk factor and the fire risk factor level. The fire risk warning value is calculated by weighting the importance of each feature obtained for several fire points in the target area. The fire risk warning value of the entire area is obtained by using the Kriging interpolation method based on this.
[0033] In step 2-2), when performing gradient boosting decision tree regression analysis to quantify the importance weights of each fire risk factor, the following steps are adopted: 1) Initialization, using the target variables in the training set. The average number of fire points is used as the initial prediction value. In the formula: Represents the total number of samples. For the first The number of fire points per sample, initializing the prediction function. As the baseline prediction function for the first iteration of the gradient boosting decision tree model, it is used for subsequent residual calculation and model update;
[0034] 2) Iterative training process, in the first... In the next iteration, the predicted value is calculated. Compared with the true value The residual (negative gradient), where when m=1, the This is the initial prediction function obtained in step 1. :
[0035] In the formula, It is the objective function. The model is for the first The prediction function for each sample. It is the first After the nth iteration, the model... For each sample, the prediction result, for the mean squared error, the residual is equivalent to the difference between the current error value and the true value:
[0036] 3) Utilizing residuals As a pseudo-target value, train a new learner. Fit the current residual distribution; combine the current learner's predictions to update the model's predictions:
[0037] In the formula: It is the first The warning function at the next iteration It is the learning rate, which controls the step size of each update to avoid overfitting. It is in the The learner, trained using the residual as a pseudo-target value in each iteration, continuously optimizes the predicted values through multiple iterations. Until the residual converges and the region of change of the objective function stabilizes; 4) Use the Bayesian optimizer to perform hyperparameter tuning (number of decision trees, learning rate, maximum depth of decision trees) to obtain the optimal operating parameters of the GBDT classifier; 5) To quantitatively assess the relative importance of each fire risk factor to the early warning model, based on the feature importance calculation mechanism built into GBDT, a feature contribution evaluation system is constructed by statistically analyzing the average information gain (measured by the reduction in mean squared error) brought by each feature when splitting all regression tree nodes. Equation (5) is used to quantify the importance weight of each feature:
[0038] in Indicates the first One characteristic, For the total number of decision trees, For the first Used in trees The set for which nodes are split. This represents the decrease in mean square error caused by the split point S.
[0039] In step 3, based on the regional-scale global fire risk warning level distribution results obtained in step 2, the fire risk warning value corresponding to each tower within the transmission line corridor is used as the dependent variable. Specifically, the regional fire risk warning value corresponding to each tower within the transmission line corridor is obtained using the mask extraction method in ArcGIS. Further, a correction factor reflecting the on-site characteristics of the transmission line is introduced as an independent variable. This correction factor includes the spatial distance between the transmission line tower and historical fire points, the attribute characteristics of being located in the inverted triangle key treatment area, and the minimum clearance distance between the transmission line tower and the ground.
[0040] Before regression analysis, spatial autocorrelation tests were performed on the dependent variable, and multicollinearity was diagnosed on the independent variable. Based on meeting the regression modeling conditions, the Geographically Weighted Regression (GWR) method was used to locally correct the global fire risk warning value at the regional scale, obtaining the corrected wildfire risk warning value for each transmission line tower. Simultaneously, by analyzing the spatial variation characteristics of the regression coefficients, the spatial heterogeneity of fire risk factors within the transmission corridor was characterized, and the fire risk factors were classified according to these spatial heterogeneity characteristics, ultimately generating a transmission corridor tower-level fire risk warning distribution map with target resolution. In step 3, the following steps were adopted: Step 3-1) Preliminary GWR Modeling: Within the target area, a preliminary geographically weighted regression (GWR) model is established for fire hazard factors and fire hazard warning risk values. The spatial coordinates of sample points are embedded with regression coefficients to obtain local regression coefficients and local fit values for each sample point, providing foundational data for subsequent spatial feature analysis. Independent variables include not only existing fire hazard factors such as meteorology, remote sensing, and topography, but also on-site factors related to transmission lines, including the minimum clearance distance between transmission line towers, the " / \" key management area markers, and the shortest distance from each tower to the fire point. The dependent variable is the fire hazard warning risk value, which originates from the global fire hazard warning map generated in Step 2 using the GBDT method, and the local risk value of each fire point is obtained after target area masking. By mapping fire hazard factors to fire hazard risk values according to spatial coordinates, the GWR model can quantify the local contribution of each factor to fire hazard risk at different geographical locations, thus providing a foundation for subsequent spatial autocorrelation analysis, multicollinearity diagnosis, and final GWR model optimization. Geographically Weighted Regression (GWR) is a regression analysis method that captures spatial heterogeneity. It allows regression coefficients to vary spatially, considers the local effects of spatial objects, and embeds the geographical location of sample data into the regression coefficients in the form of distance weights. The expression of the GWR model is shown in equation (6):
[0041] In the formula: It is the first The spatial coordinates of each sample point; Indicates the first The sample point of the th sample point Local regression coefficients of the independent variables, These are independent and identically distributed error terms. The spatial coordinates... It is used to characterize the relative positional relationship of sample points in geospace and serves as a unified spatial benchmark for subsequent spatial autocorrelation analysis, spatial weight calculation, and geographic weighted regression parameter estimation.
[0042] Step 3-2) Spatial Autocorrelation Analysis: Before performing geographically weighted regression analysis, it is necessary to assess the spatial distribution characteristics of the dependent variable, the fire risk index, and determine its spatial distribution pattern. When significant spatial clustering exists, it indicates that the dependent variable is suitable for GWR modeling, and the local regression coefficients can be used to reflect spatial heterogeneity. The Moran's index is a commonly used global spatial autocorrelation test method to measure the spatial distribution pattern of variables. A positive Moran's index indicates that the data exhibits a clustered distribution, while a negative Moran's index indicates that the data exhibits a discrete distribution. The formula for calculating the Moran's index is shown in Equation 7:
[0043] The total number of observation units, , The first The and the first Fire risk index for each unit, This represents the average of the fire risk index. The spatial weight matrix represents the unit. and unit The spatial adjacency relationship between them, wherein the spatial weight matrix Based on the sample points being in the same spatial coordinate system The spatial autocorrelation of the fire risk index was constructed using distance relationships. Further statistical tests were conducted to verify this. The Moran's index was 0.881, significantly close to 1, indicating a strong positive spatial correlation. High and low values clustered spatially. The Z-value, a standardized value of the observed Moran's index, was 14.89, far exceeding 1.96 (corresponding to a 95% confidence level). Furthermore, the p-value's significance probability was close to 0 and far less than 0.05, supporting the rejection of the null hypothesis. This indicates that the spatial clustering is highly statistically significant. The spatial distribution pattern of the fire risk index is suitable for GWR modeling to explore the spatial heterogeneity of different fire risk factors.
[0044] Step 3-3) Multicollinearity Diagnosis: After completing the spatial autocorrelation analysis in Step 3-2 and confirming the significant spatial heterogeneity of the dependent variable, fire risk warning value, it is necessary to test the multicollinearity of the independent variables in the GWR model to ensure the model's stability and explanatory power. Based on local spatial characteristics, the variance inflation factor (VIF) is used to diagnose candidate fire risk factors, eliminating independent variables with high collinearity, thereby selecting the final combination of independent variables suitable for GWR modeling. Multicollinearity Diagnosis: Before performing geographically weighted regression, multicollinearity diagnosis of the independent variables is required to eliminate variables with multicollinearity to improve model accuracy. This invention uses the variance inflation factor (VIF) diagnostic method for multicollinearity diagnosis. The expression for the variance inflation factor is:
[0045] In the formula: as independent variable The multiple correlation coefficients of linear regression analysis performed on the remaining (n-1) independent variables. When When there is no multicollinearity among the variables; when There is strong multicollinearity among the variables. In this invention, the VIF values of all candidate fire risk factors are calculated for preliminary diagnosis. For factors with VIF values significantly higher than the threshold, necessary variable elimination is performed according to the diagnostic results to ensure that the combination of independent variables finally included in the GWR model meets the requirement of low collinearity, thereby improving the reliability and explanatory power of the model. Through screening, the finally selected independent variables and their VIF values are as follows: fire point distance (1.49), " / \" key area (1.19), NDVI (1.55), slope (3.25), slope aspect (1.27), population density (1.29), land cover classification (1.28), minimum clearance distance (1.09), and wind speed (1.52). All of these are significantly lower than the collinearity threshold, indicating that the correlation between variables is low, which can ensure the fitting stability and prediction accuracy of the GWR model.
[0046] Step 3-4) Local Regression Coefficients of GWR Fire Risk Factor: After completing the multicollinearity screening in Step 3-3, the final GWR is fitted using the optimized combination of independent variables. To fully reflect the local influence of sample points, a Gaussian spatial weight function is introduced. The expression of the Gaussian spatial weight function is shown in Equation 9: Where: Indicates the first A pair of objects The weight of influence For sample points With sample points In spatial coordinates and The Euclidean distance is calculated below. The bandwidth parameter determines the search range within the domain.
[0047]
[0048] Compared to global regression models By using Gaussian spatial weighting functions, each sample point has an independent set of regression coefficients, revealing the contribution and mechanism of each fire risk factor at different geographical locations. For example... Figure 10 As shown in the regression coefficient diagram of fire hazard factors, this invention classifies and analyzes fire hazard factors for transmission lines, revealing the spatial characteristics of different types of factors: for example, factors such as temperature show a linear trend with spatial dimension; the contribution of on-site factors such as minimum clearance distance of transmission lines approaches a global regression trend; while factors such as historical fire point distance and land use type show significant spatial heterogeneity. This analysis not only reveals the mechanism of action of fire hazard factors in local space, but also provides a refined and operable decision-making basis for fire hazard early warning along transmission lines, while visualizing the local contribution, making the model results intuitive and easy to understand.
[0049] Steps 3-5) GWR bandwidth optimization: This invention uses spatial statistics tools in ArcGIS to implement GWR regression. Considering that the distribution of transmission line towers is not completely uniform, during the GWR regression process, the adaptive neighborhood type is selected with a fixed number of neighborhoods as the bandwidth range to ensure that each sample point contains sufficient neighborhood information in the local regression. The Corrected Akaike Information Criterion (AIC) is selected to automatically search for the optimal bandwidth value to achieve local optimal fitting of the model.
[0050] In step 4, based on the tower-level wildfire risk distribution results obtained in step 3, a fire risk level classification threshold is set, and the wildfire risk level of each tower along the transmission line is classified and determined. The classification and determination results are compared and verified with the actual operation and maintenance classification results of the power operation and maintenance management unit. On this basis, the tower fire risk level and the spatial information of the transmission line are integrated and displayed to generate a wildfire risk early warning map of the transmission line corridor. This reduces the situation where adjacent tower sections are classified into the same early warning level, realizes intuitive identification of high-risk sections, and improves the accuracy and efficiency of wildfire risk early warning of the transmission line corridor.
[0051] Example: The technical solution of this invention will be further explained using a research example of the Xiaogan City Xiaoshi I and II power transmission corridor: The specific steps for identifying typical features in Xiaogan area of Hubei Province are as follows: S1. Data Acquisition and Processing: Surface reflectance data from hyperspectral remote sensing images of Xiaogan area were acquired from the Natural Resources Remote Sensing Platform. After cloud masking, cirrus removal, and atmospheric correction, the B1-B8 bands were used to calculate remote sensing indices such as the Normalized Difference Vegetation Index (NDVI). Meteorological data were provided by the National Data Center for the Qinghai-Tibet Plateau (1km). 1km raster data and terrain elevation data were obtained from NASA's DEM dataset. Slope and aspect were calculated from the DEM in GEE (Google Earth Engine). Because different wildfire impact factors have different spatial scales and geographic reference systems, to ensure accuracy in wildfire risk assessment, the projected coordinate system was converted to WGS_1984_UTM_Zone_50N using the projected raster tool in ArcGIS. The resampling tool was then used to uniformly set the cell size of all raster data to (1000m, 1000m).
[0052] S2. Global early warning of fire risk in power transmission corridors: Taking 932 fire points observed by the Himawari satellite from January to May 2024 as the research object, the fire risk factor data in the above steps are superimposed on each fire point to classify the fire risk factors. For positive data (monthly average temperature, fire point radiation power, wind speed), linear classification based on quantiles (0%-10%, 10%-30%, 30%-60%, 60%-90%, 90%-100%) is used, categorized into levels 1-5. For negative data (monthly average precipitation, soil moisture), linear classification based on quantiles (0%-10%, 10%-30%, 30%-60%, 60%-90%, 90%-100%) is used, categorized into levels 5-1. For moderate indicators (normalized difference vegetation index, average population density, land use type, slope aspect, elevation), classification is based on the indicator's own characteristics. For example, slope fire risk levels are divided into five levels from level one: no slope aspect, shaded slope, sunny slope, semi-shaded slope, and semi-sunny slope. Using the combination of warning indicator levels for each fire risk factor as the independent variable, the number of all fire points under that level combination is counted as the dependent variable. Gradient boosting decision tree regression analysis based on feature selection and Bayesian parameter tuning using Python is employed to quantitatively evaluate the importance weights of each fire risk factor.
[0053] Gradient Boosting Decision Tree (GBDT): The GBDT model is trained, and Bayesian optimization is used to select the optimal parameters (n_estimators=98, learning_rate=0.0976, max_depth=9). The fire risk index distribution is output, and Kriging interpolation is used to generate a global fire risk map of Xiaogan City. 1) Initialization, using the target variables in the training set. The average number of fire points was used as the initial prediction value. In the formula: Represents the total number of samples. For the first The number of fire points per sample, initializing the prediction function. It serves as the baseline prediction function for the first iteration of the gradient boosting decision tree model, and is used for subsequent residual calculations and model updates.
[0054]
[0055] 2) Iterative training process, in the first... In the next iteration, the current model prediction value is calculated. Compared with the true value The residual (negative gradient) where, when m=1, the This is the initial prediction function obtained in step 1:
[0056] In the formula, The objective function is the residual, which, for mean squared error, is equivalent to the difference between the current error value and the true value. The model is for the first The prediction function for each sample. It is the first After the nth iteration, the model... Prediction results for each sample
[0057] 3) Utilizing residuals As a pseudo-target value, train a new learner. Fit the current residual distribution. Based on the current learner's predictions, update the model's predictions:
[0058] In the formula: It is the first The warning function at the next iteration It is the learning rate, which controls the step size of each update to avoid overfitting. It is in the The learner, trained using the residual as a pseudo-target value in each iteration, continuously optimizes the predicted values through multiple iterations. The process continues until the residual converges and the region of change of the objective function stabilizes.
[0059] 4) Bayesian parameter tuning: This study uses a Bayesian optimizer to tune hyperparameters (number of decision trees, learning rate, maximum depth of decision trees) to obtain the optimal operating parameters of the GBDT classifier.
[0060] 5) To quantitatively assess the relative importance of each fire risk factor to the early warning model, this study uses the feature importance calculation mechanism built into GBDT. By statistically analyzing the average information gain (measured by the reduction in mean squared error) brought by each feature when splitting all regression tree nodes, a feature contribution evaluation system is constructed, and the importance weight of each feature is quantified using equation (5):
[0061] in Indicates the first One characteristic, For the total number of decision trees, For the first Used in trees The set for which nodes are split. This represents the decrease in mean square error caused by the split point S.
[0062] The fire risk warning value can be represented by a weighted linear sum of the characteristic contribution of fire risk factors and the level of fire risk factors. The fire risk warning value is calculated using this formula for 932 fire points in Xiaogan City. Based on this, the risk value of the entire region is obtained by Kriging interpolation and the risk is divided into 1-5 levels.
[0063] S3. Local Correction: Taking the small grids to which the 412 towers of the Xiaoshi No. 1 and No. 2 lines belong as the research objects, the global early warning value of GBDT is used as the dependent variable, and on-site factors of the transmission line (minimum clearance distance between towers, distance of historical fire points, and "∧" characteristic area) are added. Spatial autocorrelation test is performed on the dependent variable, and multicollinearity test is performed on the independent variables. After GWR regression analysis, a visual map of the corrected transmission line corridor is formed. The changes in the risk level of specific sections in the transmission line corridor are compared and demonstrated, and refined prevention and control suggestions are given.
[0064] 1) Geographically Weighted Regression (GWR): GWR is a regression analysis method that can capture spatial heterogeneity. It allows regression coefficients to vary spatially, considers the local effects of spatial objects, and embeds the geographical location of sample data into the regression coefficients in the form of distance weights. The expression of the GWR model is shown in equation (6):
[0065] In the formula: It is the first The spatial coordinates of each sample point; Indicates the first The sample point of the th sample point Local regression coefficients of the independent variables, These are independent and identically distributed error terms, and the spatial coordinates are... It is used to characterize the relative positional relationship of sample points in geospace and serves as a unified spatial benchmark for subsequent spatial autocorrelation analysis, spatial weight calculation, and geographic weighted regression parameter estimation.
[0066] 2) Spatial Autocorrelation Analysis: Before conducting geographically weighted regression analysis, it is necessary to assess whether the dependent variable, the fire risk index, exhibits spatial autocorrelation. The Moran's index is a commonly used global spatial autocorrelation test method to measure the spatial distribution pattern of variables. A positive Moran's index indicates that the data exhibits a clustered distribution, while a negative Moran's index indicates that the data exhibits a discrete distribution. The formula for calculating the Moran's index is shown in Equation 7:
[0067] The total number of observation units, , The first The and the first Fire risk index for each unit, This represents the average of the fire risk index. The spatial weight matrix represents the unit. and unit The spatial adjacency relationship between them. Wherein, the spatial weight matrix... Based on the sample points being in the same spatial coordinate system The spatial autocorrelation of the fire risk index was further verified by constructing distance relationships and combining statistical tests. The Moran index was 0.881, which is significantly close to 1, indicating a strong positive spatial correlation. High and low values clustered spatially. The Z-value, which is the standardized value of the observed Moran index, was 14.89, which is much higher than 1.96 (corresponding to the 95% confidence level). Moreover, the significance probability of the P-value was close to 0 and much less than 0.05, supporting the rejection of the null hypothesis. This indicates that the spatial clustering has extremely high statistical significance. The spatial distribution pattern of the fire risk index is suitable for GWR modeling to explore the spatial heterogeneity of different fire risk factors.
[0068] 3) Multicollinearity Diagnosis: Before performing geographically weighted regression, it is necessary to diagnose multicollinearity in the independent variables and eliminate variables exhibiting multicollinearity to improve model accuracy. This invention uses the variance inflation factor (VIF) diagnostic method for multicollinearity diagnosis. The expression for the variance inflation factor is:
[0069] In the formula: as independent variable The multiple correlation coefficients of linear regression analysis performed on the remaining (n-1) independent variables. When When there is no multicollinearity among the variables; when There is strong multicollinearity among the variables. In this invention, the VIF value of all candidate fire hazard factors is calculated for preliminary diagnosis. For factors with VIF values significantly higher than the threshold, necessary variable removal will be performed according to the diagnosis results to ensure that the combination of independent variables finally included in the GWR model meets the requirement of low multicollinearity, thereby improving the reliability and interpretability of the model.
[0070] 4) The Gaussian spatial weight function expression is shown in Equation 10: Where: Indicates the first A pair of objects The weight of influence For sample points With sample points In spatial coordinates and The Euclidean distance is calculated below. Bandwidth determines the search range within the domain.
[0071]
[0072] Compared to global regression models This invention employs a Gaussian spatial weighting function to ensure that each sample point has an independent set of regression coefficients, revealing the contribution and mechanism of each fire hazard factor at different geographical locations. The invention utilizes spatial statistical tools in ArcGIS to implement GWR regression. Considering that the distribution of transmission line towers is not completely uniform, a fixed number of adaptive domain types is selected as the bandwidth range, and the Corrected Akake Information Criterion (AIC) is used to automatically search for the optimal bandwidth value.
[0073] S4. Results Analysis: The fire hazard warning level map of the Xiaoshi I-II circuit buffer zone was obtained. Comparison with the global GBDT fire hazard warning model revealed that the fire hazard level of some sections of the Xiaoshi I circuit decreased from level 5 to level 3 after the correction, reflecting the significant impact of local environmental characteristics on risk. The low-risk area expanded by 32%. The results showed that the R² of the corrected model increased to 0.791, and the accuracy improved by 23.4% compared to the global model.
Claims
1. A multi-factor coupled method for early warning of wildfire risks along power transmission lines, characterized in that, Includes the following steps: Step 1: Collect data on several historical wildfire hotspots in the target area, and obtain meteorological data such as monthly precipitation, monthly average temperature, soil moisture and wind speed in the area. At the same time, obtain surface reflectance data affected by hyperspectral remote sensing to calculate vegetation ecological factors and perform land use classification, and construct a standardized multi-factor fire risk factor dataset for wildfire risk analysis. Step 2: Based on the multi-factor fire risk factor dataset obtained in Step 1, the fire risk factors are classified according to their physical meaning and risk change characteristics. On this basis, a sample dataset for global wildfire risk modeling is constructed, and then the model is trained and the hyperparameters are tuned to finally obtain the regional scale global fire risk level distribution results. Step 3: Based on the regional-scale global fire risk level distribution results obtained in Step 2, obtain the fire risk distribution map of transmission corridor towers; Step 4: Based on the results of the distribution of forest fire risk at the pole level obtained in Step 3, set the threshold for classifying fire risk levels and classify the forest fire risk level of each pole along the transmission line.
2. The method according to claim 1, characterized in that, In step 1, Himawari geostationary satellite observations were used to obtain data on several historical wildfire hotspots in the target study area over a period of time. Simultaneously, meteorological data such as monthly precipitation, monthly average temperature, soil moisture, and wind speed were acquired at a designated scientific data center. Meanwhile, surface reflectance data influenced by hyperspectral remote sensing were obtained from the Natural Resources Remote Sensing Cloud Service Platform to calculate NDVI vegetation ecological factors and perform land use classification. In ArcGIS, the projection raster tool was used to convert the projection coordinate system of all factors to WGS_1984_UTM_Zone_50N. The resampling tool was used to uniformly set the cell size of all raster data, constructing a standardized multi-factor fire risk factor dataset for wildfire risk analysis.
3. The method according to claim 2, characterized in that, In step 1, vegetation ecological factors include Normalized Difference Vegetation Index (NDVI), Free Fiber Cover (FVC), Digital Elevation Model (DEM), slope, and aspect. The characteristic factors of the power transmission corridor include the minimum clearance distance between transmission line towers, the inverted triangle key treatment area, the distance of historical fire points, and the distribution of cemeteries.
4. The method according to claim 1, characterized in that, In step 2, the process for conducting a global wildfire risk early warning analysis is as follows: Step 2-1) Using several fire points in the target month as targets, overlay multiple fire risk factor data onto each fire point and classify the fire risk factors. Step 2-2) Using the combination of warning index levels of each fire risk factor as the independent variable, the number of all fire points under the combination of levels is counted as the dependent variable. Using Python, gradient boosting decision tree regression analysis based on feature selection and Bayesian parameter tuning is used to quantitatively evaluate the importance weight of each fire risk factor. Steps 2-3) Obtain the fire risk warning value. The fire risk warning value is a weighted linear representation of the characteristic contribution degree of the fire risk factor and the fire risk factor level. The fire risk warning value is calculated by weighting the importance of each feature obtained for several fire points in the target area. The fire risk warning value of the entire area is obtained by using the Kriging interpolation method based on this.
5. The method according to claim 4, characterized in that, In step 2-2), when performing gradient boosting decision tree regression analysis to quantify the importance weights of each fire risk factor, the following steps are adopted: 1) Initialization, using the target variables in the training set. The average number of fire points is used as the initial prediction value. In the formula: Represents the total number of samples. For the first The number of fire points per sample, initializing the prediction function. As the baseline prediction function for the first iteration of the gradient boosting decision tree model, it is used for subsequent residual calculation and model update; 2) Iterative training process, in the first... In the next iteration, the predicted value is calculated. Compared with the true value The residual, where when m=1, the This is the initial prediction function obtained in step 1. : In the formula, It is the objective function. The model is for the first The prediction function for each sample. It is the first After the nth iteration, the model... For each sample, the prediction result, for the mean squared error, the residual is equivalent to the difference between the current error value and the true value: 3) Utilizing residuals As a pseudo-target value, train a new learner. Fit the current residual distribution; combine the current learner's predictions to update the model's predictions: In the formula: It is the first The warning function at the next iteration It is the learning rate, which controls the step size of each update to avoid overfitting. It is in the The learner, trained using the residual as a pseudo-target value in each iteration, continuously optimizes the predicted values through multiple iterations. Until the residual converges and the region of change of the objective function stabilizes; 4) Use the Bayesian optimizer to perform hyperparameter tuning to obtain the optimal operating parameters for the GBDT classifier; 5) To quantitatively assess the relative importance of each fire risk factor to the early warning model, based on the feature importance calculation mechanism built into GBDT, a feature contribution evaluation system is constructed by statistically analyzing the average information gain brought by each feature when splitting at all regression tree nodes. Equation (5) is used to quantify the importance weight of each feature: in Indicates the first One characteristic, For the total number of decision trees, For the first Used in trees The set for which nodes are split. This represents the decrease in mean square error caused by the split point S.
6. The method according to any one of claims 1 to 5, characterized in that, In step 3, the following steps are performed: Step 3-1) Perform preliminary GWR modeling; within the target area, establish a preliminary geographically weighted regression GWR model for fire risk factors and fire risk warning values, embed the regression coefficients into the spatial coordinates of the sample points to obtain the local regression coefficients and local fitting values of each sample point, providing basic data for subsequent spatial feature analysis; Step 3-2) Perform spatial autocorrelation analysis; Step 3-3) Perform multicollinearity diagnosis; After completing the spatial autocorrelation analysis in Step 3-2 and confirming that the dependent variable fire risk warning value has significant spatial heterogeneity, perform multicollinearity test on the independent variables of the GWR model to ensure the stability and explanatory power of the model. Step 3-4) Obtain the local regression coefficients of the GWR fire risk factor; specifically, after completing the multicollinearity screening in step 3-3, use the optimized combination of independent variables to perform the final fitting of GWR.
7. The method according to claim 6, characterized in that, In step 3-1), when performing preliminary GWR modeling, the independent variables not only include the original fire risk factors such as meteorology, remote sensing and topography, but also introduce on-site factors of transmission lines, including the minimum clearance distance of transmission line towers, the inverted triangle key treatment area and the shortest distance from each tower to the fire point. The dependent variable is the fire risk warning value, which is derived from the global fire risk warning map generated by the GBDT method in step 2, and the local risk value of each fire point is obtained by target area masking. By mapping fire risk factors to fire risk values using spatial coordinates, the GWR model can quantify the local contribution of each factor to fire risk at different geographical locations, thus providing a foundation for subsequent spatial autocorrelation analysis, multicollinearity diagnosis, and final GWR model optimization; the expression of the GWR model is shown in equation (6): In the formula: It is the first Spatial coordinates of each sample point Located in spatial coordinates The local intercept term at the location is used to characterize the baseline fire risk level at that location without considering the influence of various fire risk factors; Indicates the first The sample point of the th sample point Local regression coefficients of the independent variables, These are independent and identically distributed error terms; the spatial coordinates It is used to characterize the relative positional relationship of sample points in geospace and serves as a unified spatial benchmark for subsequent spatial autocorrelation analysis, spatial weight calculation, and geographic weighted regression parameter estimation.
8. The method according to claim 6, characterized in that, In step 3-2), the following formula is used for judgment when performing spatial autocorrelation analysis: The total number of observation units, , The first The and the first Fire risk index for each unit, This represents the average of the fire risk index. The spatial weight matrix represents the unit. and unit The spatial adjacency relationship between them, wherein the spatial weight matrix Based on the sample points being in the same spatial coordinate system The distance relationship is constructed below.
9. The method according to claim 6, characterized in that, In step 3-3, the variance inflation factor (VIF) diagnostic method is used to diagnose multicollinearity. The expression for the variance inflation factor is: In the formula: as independent variable The multiple correlation coefficients for linear regression analysis of the remaining (n-1) independent variables; when When there is no multicollinearity among the variables; when There is strong multicollinearity among the variables; the VIF value of all candidate fire risk factors is calculated for preliminary diagnosis. For factors with VIF values significantly higher than the threshold, necessary variable removal will be carried out according to the diagnosis results to ensure that the combination of independent variables finally included in the GWR model meets the requirement of low collinearity, thereby improving the reliability and explanatory power of the model. In steps 3-4, a Gaussian spatial weighting function is introduced to reveal the contribution and mechanism of each fire risk factor at different geographical locations; the expression of the Gaussian spatial weighting function is shown in Equation 9: In the formula: Indicates the first A pair of objects The weight of influence For sample points With sample points In spatial coordinates and The Euclidean distance is calculated below. The bandwidth parameter determines the search range within the domain.
10. The method according to any one of claims 7 to 9, characterized in that, It also includes steps 3-5), optimizing the GWR bandwidth; using spatial statistics tools in ArcGIS to perform GWR regression, considering that the distribution of transmission line towers is not completely uniform, in the GWR regression process, the adaptive neighborhood type is selected and a fixed number of neighborhoods is used as the bandwidth range to ensure that each sample point contains sufficient neighborhood information in the local regression, and the Akaike Information Criterion (AIC) correction value is selected to automatically search for the optimal bandwidth value to achieve local optimal fitting of the model.