Power transmission line forest fire monitoring and early warning method based on meteorology and three-dimensional visual support

Abstract

The invention relates to the technical field of power monitoring, in particular to a power transmission line forest fire monitoring and early warning method based on meteorology and three-dimensional visual support, which comprises the following steps: S1, acquiring multi-source data of a target area and preprocessing the multi-source data; s2, predicting the fire risk of the target area by adopting a TFT model in combination with multi-source data; s3, constructing a visual three-dimensional geographic model; when the predicted fire risk level exceeds a preset safety threshold value, early warning is triggered, fire early warning information is pushed and displayed in the three-dimensional geographic model, and the fire early warning information comprises geographic boundary coordinates, the fire risk level and electric power facility information. The fire risk can be monitored in real time, the fire occurrence probability is analyzed in combination with meteorological data, and the fire development trend and the risk area are displayed through a three-dimensional visual interface.

Description

Technical Field

[0001] The present invention relates to the technical field of electric power monitoring, and in particular to a method for monitoring and early warning of forest fires on power transmission lines based on meteorological and three-dimensional visualization support. Background Art

[0002] In recent years, frequent wildfires have become one of the major disasters threatening the safety of transmission lines and the stable operation of power grids, especially in Gansu, which has a complex terrain, drought and little rain. Its impact on transmission lines is becoming increasingly significant. Wildfires may not only cause transmission line tripping and high-voltage insulation damage, but also have a serious impact on equipment stability, leading to power outages and safety hazards.

[0003] Current wildfire early warning methods mainly rely on meteorological data and satellite monitoring, but in areas with complex terrain, how to conduct refined early warning analysis remains a challenge, especially the limited ability to effectively monitor small-area fires. In addition, the existing early warning methods have a weak correlation with power grid equipment, making it difficult to provide accurate disaster warnings and timely emergency support. Therefore, developing a method based on meteorological data, combined with high-precision geographic information, and using three-dimensional visualization to achieve wildfire monitoring and early warning has important practical significance and application prospects. Summary of the invention

[0004] In order to solve the above problems, the present invention provides a transmission line wildfire monitoring and early warning method based on meteorological and three-dimensional visualization support, which can monitor fire risks in real time, analyze the probability of fire occurrence in combination with meteorological data, and display fire development trends and risk areas through a three-dimensional visualization interface.

[0005] The technical solution adopted by the present invention is to provide a transmission line wildfire monitoring and early warning method based on meteorological and three-dimensional visualization support, comprising the following steps:

[0006] S1. Obtain and pre-process multi-source data of the target area, including meteorological data, historical fire data, environmental data, geographic data, and power facility information;

[0007] S2, using the TFT model combined with multi-source data to predict the fire risk of the target area;

[0008] S3, build a visual three-dimensional geographic model to display current meteorological data, environmental data and geographic data in real time;

[0009] When the predicted fire risk level exceeds the preset safety threshold, an early warning is triggered, and fire warning information is pushed and displayed on the three-dimensional geographic model. The fire warning information includes geographic boundary coordinates, fire risk level and power facility information.

[0010] The preprocessing of multi-source data in step S1 includes data cleaning and standardization.

[0011] The data cleaning is to screen the multi-source data using a threshold detection method within a preset range and delete abnormal data;

[0012] The standardization process converts the units and formats of multi-source data through a unit conversion function, parses the original data fields of the multi-source data through a parser function, extracts the data fields nested in the original data structure, and maps the extracted data fields to a unified and standardized target field through a field specification mapping table.

[0013] The meteorological data include real-time and forecast temperature, humidity, atmospheric pressure, wind speed, wind force, wind direction and rainfall.

[0014] The prediction method of the TFT model in step S2 comprises the following steps:

[0015] S201, receiving multi-source data, generating time step information, and defining fire risk levels;

[0016] S202, using a multivariate correlation attention mechanism to generate a feature matrix, specifically using historical fire data and meteorological data as input features, and calculating the interactive weights of meteorological factors contained in the meteorological data based on the head attention mechanism and the interactive attention mechanism, capturing the dynamic interactions between meteorological factors, and generating a dynamic correlation feature matrix;

[0017] S203, using a gated loop mechanism to generate a time series matrix, specifically, obtaining a feature vector according to the correlation feature matrix as an input feature of the gated loop mechanism, and adjusting the weight ratio between meteorological factors according to the change trend of the input feature in a dynamic weighted manner to generate a time series matrix with multiple time steps;

[0018] S203, fusing the time series matrix, time step information, and correlation feature matrix, and combining environmental data, geographic data, and power facility information to generate multi-dimensional structured data, including fire risk levels, fire probability, and geographic boundary coordinates of high-risk fire areas.

[0019] The dynamic weighting in step S203 is specifically to add a dynamic weighting term for the input change rate on the basis of the GRU update formula, wherein the GRU update formula is:

[0020] Update Gate:

[0021] z t =σ(W z x t +U z h t-1+b z )

[0022] Where t represents the time step, z t represents the output of the update gate, σ represents the sigmoid activation function, x t represents the input of the current time step, h t-1 represents the hidden state of the previous time step, W z and U z represents the weight matrix, b z represents the bias term;

[0023] Reset the gate:

[0024] r t =σ(W r x t +U r h t-1 +b r )

[0025] Among them, t represents the time step, r t represents the output of the reset gate, σ represents the sigmoid activation function, x t represents the input of the current time step, h t-1 represents the hidden state of the previous time step, W r and U r represents the weight matrix, b r represents the bias term;

[0026] Candidate hidden states:

[0027]

[0028] Where t represents the time step, represents the output of the candidate hidden state, tanh represents the hyperbolic tangent activation function, x t represents the input of the current time step, h t-1 represents the hidden state of the previous time step, W h and U h represents the weight matrix, b h represents the bias term; ⊙ represents element-by-element multiplication.

[0029] Final hidden state:

[0030]

[0031] Where t represents the time step, h t represents the final hidden state output, h t-1 represents the hidden state of the previous time step, and ⊙ represents element-wise multiplication.

[0032] On this basis, the time-sensitive weighting factor Δ is introduced t , Δ t The calculation formula is,

[0033]

[0034] Δ t Represents the rate of change of input data, ||x t -x t-1 || represents the change range between the current input and the previous input, α represents the adjustment factor, which is used to control the influence of the change rate on the dynamic change of the weight, and ∈ represents a small value, usually 10 -6 or 10 -8 , to avoid the denominator being zero, max is the maximum value of the number of elements in the formula;

[0035] Introducing Δ t Dynamically adjusted update gate:

[0036] z′ t =z t ·(1+β·Δ t )

[0037] Where t represents the time step, z′ t Represents the introduction of Δ t The output of the dynamically adjusted update gate, β is the dynamic adjustment coefficient, which controls the weight amplification effect of input changes on the update gate, z t Represents the output of the original update gate;

[0038] Introducing Δ t Dynamically adjusted hidden state:

[0039]

[0040] Where t represents the time step, h′ t Represents the introduction of Δ t The output of the dynamically adjusted hidden state, h t-1 represents the hidden state of the previous time step, and ⊙ represents element-wise multiplication.

[0041] The beneficial effect of the present invention is that it provides a transmission line wildfire monitoring and early warning method based on meteorological and three-dimensional visualization support, and forms an improved TFT model by embedding a customized attention mechanism and an improved gated loop mechanism in the TFT model, which can better process multi-source meteorological data, accurately capture the dynamic relationship between various meteorological factors, and provide dynamic fire risk prediction through time series modeling. The customized attention mechanism and the gated loop mechanism work together in a sequential processing manner to ensure that the model can effectively utilize the fusion of historical data and real-time data, improve the accuracy and response speed of fire risk prediction, construct a three-dimensional geographic scene of the target area, associate it with the improved TFT model, and display the meteorological data and early warning information of the target area in real time. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0043] like Figure 1 As shown, a transmission line wildfire monitoring and early warning method based on meteorological and three-dimensional visualization support includes the following steps:

[0044] S1. Obtain and pre-process multi-source data of the target area, including meteorological data, historical fire data, environmental data, geographic data, and power facility information;

[0045] S2, using the TFT model combined with multi-source data to predict the fire risk of the target area;

[0046] S3, build a visual three-dimensional geographic model to display current meteorological data, environmental data and geographic data in real time;

[0047] When the predicted fire risk level exceeds the preset safety threshold, an early warning is triggered, and fire warning information is pushed and displayed on the three-dimensional geographic model. The fire warning information includes geographic boundary coordinates, fire risk level and power facility information.

[0048] like Figure 1 As shown, the preprocessing of multi-source data in step S1 includes data cleaning and standardization.

[0049] like Figure 1 As shown, the data cleaning is to screen multi-source data using a threshold detection method within a preset range and delete abnormal data;

[0050] like Figure 1As shown, the standardization processing is to convert the units and formats of multi-source data through a unit conversion function, parse the original data fields of the multi-source data through a parser function, extract the data fields nested in the original data structure, and map the extracted data fields to a unified and standardized target field through a field specification mapping table.

[0051] like Figure 1 As shown, the meteorological data include real-time and forecast temperature, humidity, atmospheric pressure, wind speed, wind force, wind direction and rainfall.

[0052] like Figure 1 As shown, the prediction method of the TFT model in step S2 includes the following steps:

[0053] S201, receiving multi-source data, generating time step information, and defining fire risk levels;

[0054] S202, using a multivariate correlation attention mechanism to generate a feature matrix, specifically using historical fire data and meteorological data as input features, and calculating the interactive weights of meteorological factors contained in the meteorological data based on the head attention mechanism and the interactive attention mechanism, capturing the dynamic interactions between meteorological factors, and generating a dynamic correlation feature matrix;

[0055] S203, using a gated loop mechanism to generate a time series matrix, specifically, obtaining a feature vector according to the correlation feature matrix as an input feature of the gated loop mechanism, and adjusting the weight ratio between meteorological factors according to the change trend of the input feature in a dynamic weighted manner to generate a time series matrix with multiple time steps;

[0056] S203, fusing the time series matrix, time step information, and correlation feature matrix, and combining environmental data, geographic data, and power facility information to generate multi-dimensional structured data, including fire risk levels, fire probability, and geographic boundary coordinates of high-risk fire areas.

[0057] like Figure 1 As shown, the dynamic weighting in step S203 is specifically to add a dynamic weighting term for the input change rate on the basis of the GRU update formula, wherein the GRU update formula is:

[0058] Update Gate:

[0059] z t =σ(W z x t +U z h t-1 +b z )

[0060] Where t represents the time step, z trepresents the output of the update gate, σ represents the sigmoid activation function, x t represents the input of the current time step, h t-1 represents the hidden state of the previous time step, W z and U z represents the weight matrix, b z represents the bias term;

[0061] Reset the gate:

[0062] r t =σ(W r x t +U r h t-1 +b r )

[0063] Among them, t represents the time step, r t represents the output of the reset gate, σ represents the sigmoid activation function, x t represents the input of the current time step, h t-1 represents the hidden state of the previous time step, W r and U r represents the weight matrix, b r represents the bias term;

[0064] Candidate hidden states:

[0065]

[0066] Where t represents the time step, represents the output of the candidate hidden state, tanh represents the hyperbolic tangent activation function, x t represents the input of the current time step, h t-1 represents the hidden state of the previous time step, W h and U h represents the weight matrix, b h represents the bias term; ⊙ represents element-by-element multiplication.

[0067] Final hidden state:

[0068]

[0069] Where t represents the time step, h t represents the final hidden state output, h t-1 represents the hidden state of the previous time step, and ⊙ represents element-wise multiplication.

[0070] On this basis, the time-sensitive weighting factor Δ is introduced t , Δ t The calculation formula is,

[0071]

[0072] Δ t Represents the rate of change of input data, ||x t -x t-1 || represents the change range between the current input and the previous input, α represents the adjustment factor, which is used to control the influence of the change rate on the dynamic change of the weight, and ∈ represents a small value, usually 10 -6 or 10 -8 , to avoid the denominator being zero, max is the maximum value of the number of elements in the formula;

[0073] Introducing Δ t Dynamically adjusted update gate:

[0074] z′ t =z t ·(1+β·Δ t )

[0075] Among them, t represents the time step, z′ represents the introduction of Δ t The output of the dynamically adjusted update gate, β is the dynamic adjustment coefficient, which controls the weight amplification effect of input changes on the update gate, z t Represents the output of the original update gate;

[0076] Introducing Δ t Dynamically adjusted hidden state:

[0077]

[0078] Where t represents the time step, h′ t Represents the introduction of Δ t The output of the dynamically adjusted hidden state, h t-1 represents the hidden state of the previous time step, and ⊙ represents element-wise multiplication.

[0079] The present invention can realize multi-source data integration and analysis, utilize historical fire data and real-time meteorological information, and intelligently generate fire warnings, thereby providing reliable data support and decision-making basis for the protection of power facilities and the prevention of wildfire risks.

[0080] In the comprehensive analysis and early warning of fire risks, the Temporal Fusion Transformer large model, hereinafter referred to as the TFT model, is used. Based on the TFT model, a customized multivariate correlation attention mechanism and an improved gated loop mechanism are introduced to form an improved TFT model. The improved TFT model can more accurately process the dynamic characteristics and temporal dependencies of the input data, and finally generate structured prediction data. The following is the specific process of the improved TFT model to achieve structured data output:

[0081] 1. Preprocessing and feature vector of input data:

[0082] First, multi-source data is received and standardized and normalized to ensure that each variable is in the same scale range, which is conducive to subsequent calculations. Subsequently, the improved TFT model generates time step information, such as season and time period, based on multi-source data to provide time series features. On this basis, a customized attention mechanism is introduced to combine multi-head self-attention and interactive attention to calculate the interaction weights between meteorological factors, generate a correlation feature matrix, and capture the dynamic interactions between meteorological variables. The attention mechanism dynamically adjusts the weights of each meteorological variable to improve the accuracy of fire risk assessment. Finally, these processed weighted multi-dimensional dynamic feature matrix vectors are passed to subsequent calculations, retaining the relationship between meteorological data and combining with fire risk correlation to provide optimized input for subsequent time series modeling.

[0083] 2. Time series feature extraction and time series status update

[0084] Entering the time series modeling stage, the feature vector processed by the customized attention mechanism is passed to the improved gated recurrent unit to further optimize the time series feature extraction. The feature vector is calculated by the correlation feature matrix. At this stage, the improved TFT model replaces the native GRU / LSTM module of the TFT model by introducing multiple weight matrices and gated vector combinations, so that it can adaptively adjust the weighted relationship between historical fire data and real-time meteorological changes. This mechanism adjusts the data fusion ratio through a dynamic weighted formula to accurately capture the time series characteristics related to fire risk. At the same time, the time series state is dynamically weighted and updated according to the changing trend of the real-time input data, so that the TFT model can better reflect the dynamic changes of fire risk, especially when the meteorological conditions change rapidly. Make accurate predictions. Finally, these processing results generate dynamic feature matrices of multiple time steps, each time step corresponds to a complete time series risk state, forming a time series matrix.

[0085] 3. Data integration and structured output

[0086] The generated time series matrix, time step information, and correlation feature matrix output by the customized attention mechanism are fused for structured output. First, the time series feature matrix is ​​combined with the spatial encoding information of the fire hazard area to form a richer feature input. Then, these features are aggregated through the fully connected layer and mapped to specific output categories to generate multi-dimensional structured data, including fire risk level, which is defined by historical fire data, such as low risk, medium risk and high risk; the probability of fire occurrence, which is between [0,1], indicating the probability of fire occurrence, and the geographic boundary coordinates of high-risk fire areas, indicating the latitude and longitude range of the fire area, for geospatial positioning. Finally, these output data are structured in the form of tensors, covering all dimensions of fire risk assessment, and simply expressing information including risk level, probability of occurrence, and regional boundary coordinates.

[0087] Regarding the display of structured data: The improved TFT model works together through a customized attention mechanism and gated recurrent units to complete feature fusion, time series modeling, and risk level classification, and generates structured output data such as fire risk level, occurrence probability, and geographic coordinates, which are usually represented in the form of tensors.

[0088] Although the improved TFT model can complete data integration and structured output, in some application scenarios, it may be necessary to cooperate with other programs for further optimization and presentation of results. For example: 1. Data formatting, converting tensors into formats such as JSON, CSV or database tables; 2. Result integration, integrating the prediction results of multiple tasks into a unified report; 3. Visualization, using GIS tools to present fire risk levels and geographic coordinates on a map to help decision makers conduct spatial risk assessment.

[0089] About multi-source data integration: In order to achieve accurate fire monitoring and early warning, it is necessary to integrate multiple key data sources. Integrate meteorological data from the National Meteorological Observatory or other authoritative meteorological agencies, including basic meteorological elements such as temperature, humidity, precipitation, and meteorological forecast data, to evaluate the impact of environmental conditions on the occurrence of fires. Collect and integrate historical fire data, such as records from fire departments, power companies, and public fire case databases, to analyze the temporal and spatial distribution patterns of fires. Obtain power company infrastructure information, including the geographical distribution of substations, transmission lines, and distribution equipment, the line channel environment, and its operating status, so as to evaluate the potential risk impact of power facilities on fires. Build a digital terrain model of the target area, integrate geographic information data such as land use type, vegetation cover, and slope, and provide spatial reference for fire prediction models. The fusion of these multi-source data will provide comprehensive and powerful support for accurate analysis and early warning of fires.

[0090] About real-time data collection: In order to improve the accuracy of fire monitoring and early warning, satellite remote sensing technology is used to regularly collect a wide range of environmental data, such as vegetation coverage, fire heat source monitoring, and surface temperature changes. These data have good timeliness and spatial resolution, and provide an important reference for fire risk early warning. In key areas, drones are used for low-altitude patrol monitoring to collect on-site meteorological data and environmental information, especially in remote areas that are sparsely populated or difficult to reach. UAV inspections can effectively supplement the gaps in ground monitoring. IoT sensor networks, such as temperature and humidity sensors and other status monitoring equipment, are deployed in target areas to build a real-time environmental monitoring system to ensure that meteorological and environmental data can be accurately obtained under different terrain and local environmental conditions.

[0091] Regarding data preprocessing: Since data from different sources may have inconsistent formats, missing data or noise problems, the collected multi-source data needs to be cleaned and standardized.

[0092] Through the combination of parser functions and mapping tables, the original data fields from different sources are parsed and mapped to unified and standardized target fields. Specifically, the parser function uses jsonpath or XPath technology to extract fields in nested data structures, while the mapping table standardizes data fields through the field specification mapping table to ensure data field consistency.

[0093] Through custom conversion rules, we use unit conversion functions to perform unified conversions for inconsistent units and formats of different data sources. Taking temperature and precipitation as examples, temperature data is converted from Fahrenheit to Celsius, and precipitation is converted from inches to millimeters. This ensures the uniformity of the physical meaning of the input data.

[0094] The threshold detection method with a preset range is used to eliminate abnormal data to ensure the accuracy and stability of the data. This detection process not only effectively removes noise data, but also improves the reliability of model data through multi-level cleaning, laying a solid data foundation for model training and prediction.

[0095] Through the above preprocessing, the TFT model can be compatible with and associated with multi-source data, and provide good data support for further fire risk analysis, thereby improving the automation level of the entire data processing process and the innovation of data preprocessing.

[0096] Customized attention mechanism for multivariate association: The customized multivariate association attention mechanism is inserted into the input layer of the TFT model, which is mainly responsible for processing multi-source meteorological data, such as temperature, humidity, wind speed, etc. The traditional TFT attention mechanism identifies the relationship between different variables by weighting the input features, but for fire risk assessment, the dynamic interaction between meteorological variables has more complex changing characteristics. Therefore, the customized attention mechanism can capture the interactive relationship between these variables in real time and automatically adjust the weight of each variable by introducing a dynamic adjustment mechanism.

[0097] The customized attention mechanism uses multi-head self-attention and interactive attention methods to calculate the weights between different meteorological factors, such as temperature, humidity, wind speed, etc., and dynamically updates these weights to generate a feature matrix with higher correlation. In this way, the TFT model can dynamically adjust the influence of variables at different time steps to ensure that the assessment of fire risk can fully consider the impact of each meteorological variable on the probability of fire. Finally, the processed feature matrix will be used as the input of the middle layer of the model.

[0098] About the improved gated recurrent mechanism: The original time series feature extraction unit of the TFT model has been replaced with an improved gated recurrent mechanism. The TFT model originally used standard GRU or LSTM to process time series data, but such models can only statically capture the relationship between historical data and current data, and cannot effectively and adaptively adjust the weighting of input data. Especially when dealing with events such as fires that are very sensitive to environmental changes, the existing gating mechanism has certain limitations in dealing with dynamic changes.

[0099] In order to solve this problem, the present invention designs an improved version of the gated recurrent unit, which adds multiple weight matrices and gated vector combinations on the basis of the traditional GRU. This allows the model to dynamically adjust the weights of different time steps according to the changes in input data and historical states, so that the temporal dependencies between historical fire data and real-time meteorological changes can be more accurately captured when processing time series data. This improved gating mechanism not only enhances the ability to extract time series features of fire risk, but also improves the model's ability to respond to dynamic environmental changes. In this way, the model can better reflect the changes in fire risk over time, especially for dynamic prediction under rapidly changing meteorological conditions.

[0100] The improvements based on the traditional GRU are specifically reflected in the following aspects:

[0101] (1) Design of multiple weight matrices: An independent weight matrix is ​​designed for each input feature. For example, different weight matrices are used for real-time weather data and historical fire data, thereby achieving differentiated processing of different data sources. At each time step, the weight matrix is ​​dynamically updated according to the input features and historical status to enhance the ability to capture key features. For example, when the real-time weather changes dramatically, the weight of the real-time data will increase, and the weight of the historical data will decrease accordingly.

[0102] (2) Optimization of gating vector combination: The update gate and reset gate in the gating mechanism introduce a multi-channel vector combination strategy to encode the features of different time steps separately and capture the dynamic changes within a specific time range. The improved gating vector not only considers the input of the current time step, but also combines the future forecast trend, which can be achieved through the meteorological data predicted in advance, thus improving the model's ability to adapt to rapidly changing environments.

[0103] (3) Dynamic weighted update formula:

[0104] The improved mechanism introduces a time-sensitive weight update formula to adaptively adjust the weighting of time series features according to the changing trend of real-time input data. For example, by additionally modeling the rate of change of meteorological data, the proportion of current state updates is dynamically adjusted. The updated state is more in line with the drastic changes in fire scenarios.

[0105] Improve the time-sensitive weight update formula in the mechanism. This formula is based on the standard GRU update formula, adding a dynamic weighting term for the input change rate, such as the change rate of meteorological data, so that the model can adapt to drastically changing scenarios.

[0106] The update formula of the standard GRU is as follows,

[0107] Update Gate:

[0108] z t =σ(W z x t +U z h t-1 +b z )

[0109] Where t represents the time step, z t represents the output of the update gate, σ represents the sigmoid activation function, x t represents the input of the current time step, h t-1 represents the hidden state of the previous time step, W z and U z represents the weight matrix, b z represents the bias term;

[0110] Reset the gate:

[0111] r t =σ(W r x t +U r h t-1 +b r )

[0112] Among them, t represents the time step, r t represents the output of the reset gate, σ represents the sigmoid activation function, x t represents the input of the current time step, h t-1 represents the hidden state of the previous time step, W r and U r represents the weight matrix, b r represents the bias term;

[0113] Candidate hidden states:

[0114]

[0115] Where t represents the time step, represents the output of the candidate hidden state, tanh represents the hyperbolic tangent activation function, x t represents the input of the current time step, h t-1 represents the hidden state of the previous time step, W h and U h represents the weight matrix, b h represents the bias term; ⊙ represents element-by-element multiplication.

[0116] Final hidden state:

[0117]

[0118] Where t represents the time step, h t represents the final hidden state output, h t-1 represents the hidden state of the previous time step, and ⊙ represents element-wise multiplication.

[0119] On this basis, the time-sensitive weighting factor Δ is introduced t , Δ t The calculation formula is,

[0120]

[0121] Δ t Represents the rate of change of input data, ||x t -x t-1 || represents the change range between the current input and the previous input, α represents the adjustment factor, which is used to control the influence of the change rate on the dynamic change of the weight, and ∈ represents a small value, usually 10 -6or 10 -8 , to avoid the denominator being zero, max is the maximum value of the number of elements in the formula;

[0122] Introducing Δ t Dynamically adjusted update gate:

[0123] z′ t =z t ·(1+β·Δ t )

[0124] Where t represents the time step, z′ t Represents the introduction of Δ t The output of the dynamically adjusted update gate, β is the dynamic adjustment coefficient, which controls the weight amplification effect of input changes on the update gate, z t Represents the output of the original update gate;

[0125] Introducing Δ t Dynamically adjusted hidden state:

[0126]

[0127] Where t represents the time step, h′ t Represents the introduction of Δ t The output of the dynamically adjusted hidden state, h t-1 represents the hidden state of the previous time step, and ⊙ represents element-wise multiplication.

[0128] Dynamically adjust the gating weight by Δ t , the model can perceive drastic changes in input data, such as sudden changes in meteorological conditions, and appropriately amplify the weight of the update gate so that the hidden state depends more on the current input rather than the historical state. Adapting to dynamic environments, this improvement is particularly useful in fire scenarios, because sudden changes in meteorological conditions may greatly affect the fire risk, and the model needs to adjust faster to adapt to new input data. And the values ​​of α and β are adjusted according to the scenario to balance the sensitivity of input changes with the stability of the model.

Claims

1. A method for monitoring and early warning of forest fires on power transmission lines based on meteorological and three-dimensional visualization support, characterized in that: The following steps are included: S1. Obtain and pre-process multi-source data of the target area, where the multi-source data includes meteorological data, historical fire data, environmental data, geographic data, and power facility information; S2, using the TFT model combined with multi-source data to predict the fire risk of the target area and generate fire prediction information; S3, build a visual three-dimensional geographic model to display current meteorological data, environmental data and geographic data in real time; When the predicted fire risk level exceeds the preset safety threshold, an early warning is triggered, and fire warning information is pushed and displayed on the three-dimensional geographic model. The fire warning information includes geographic boundary coordinates, fire risk level and power facility information.

2. According to claim 1, a method for monitoring and early warning of power transmission line wildfires based on meteorological and three-dimensional visualization support is characterized in that: The preprocessing of multi-source data in step S1 includes data cleaning and standardization.

3. According to claim 2, a method for monitoring and early warning of power transmission line wildfires based on meteorological and three-dimensional visualization support is characterized in that: The data cleaning process includes screening multi-source data using a threshold detection method within a preset range and deleting abnormal data.

4. The method for monitoring and early warning of power transmission line wildfires based on meteorological and three-dimensional visualization support according to claim 2 is characterized in that: The standardization processing includes converting the units and formats of multi-source data through a unit conversion function, parsing the original data fields of the multi-source data through a parser function, extracting the data fields nested in the original data structure, and mapping the extracted data fields to a unified and standardized target field through a field specification mapping table.

5. The method for monitoring and early warning of power transmission line wildfires based on meteorological and three-dimensional visualization support according to claim 1 is characterized in that: The meteorological data includes real-time and forecasted temperature, humidity, atmospheric pressure, wind speed, wind force, wind direction and rainfall.

6. The method for monitoring and early warning of power transmission line wildfires based on meteorological and three-dimensional visualization support according to claim 1, characterized in that: Prediction method of TFT model in S2 step The following steps are included: S201, receiving multi-source data, generating time step information, and defining fire risk levels; S202, using a multivariate correlation attention mechanism to generate a feature matrix, specifically using historical fire data and meteorological data as input features, and based on a multi-head attention mechanism and an interactive attention mechanism, performing interactive weight calculations on meteorological factors contained in the meteorological data, capturing dynamic interactions between meteorological factors, and generating a dynamic correlation feature matrix; S203, using a gated loop mechanism to generate a time series matrix, specifically, obtaining a feature vector according to the correlation feature matrix as an input feature of the gated loop mechanism, and adjusting the weight ratio between meteorological factors according to the change trend of the input feature in a dynamic weighted manner to generate a dynamic time series matrix of multiple time steps; S203, fusing the time series matrix, time step information, and correlation feature matrix, and combining environmental data, geographic data, and power facility information to generate multi-dimensional structured data, including fire risk levels, fire probability, and geographic boundary coordinates of high-risk fire areas.

7. The method for monitoring and early warning of power transmission line wildfires based on meteorological and three-dimensional visualization support according to claim 6 is characterized in that: The dynamic weighting in step S203 includes adding a dynamic weighting term for the input change rate based on the GRU update formula, where the GRU update formula is: Update Gate: z t =σ(W z x t +U z h t-1 +b z ) Where t represents the time step, z t represents the output of the update gate, σ represents the sigmoid activation function, x t represents the input of the current time step, h t-1 represents the hidden state of the previous time step, W z and U z represents the weight matrix, b z represents the bias term; Reset the gate: r t =σ(W r x t +U r h t-1 +b r ) Among them, t represents the time step, r t represents the output of the reset gate, σ represents the sigmoid activation function, x t represents the input of the current time step, h t-1 represents the hidden state of the previous time step, W r and U r represents the weight matrix, b r represents the bias term; Candidate hidden states: Where t represents the time step, represents the output of the candidate hidden state, tanh represents the hyperbolic tangent activation function, x t represents the input of the current time step, h t-1 represents the hidden state of the previous time step, W h and U h represents the weight matrix, b h represents the bias term; ⊙ represents element-by-element multiplication. Final hidden state: Where t represents the time step, h t represents the final hidden state output, h t-1 represents the hidden state of the previous time step, and ⊙ represents element-wise multiplication. On this basis, the time-sensitive weighting factor Δ is introduced t , Δ t The calculation formula is, Δ t Represents the rate of change of input data, ||x t -x t-1 || represents the change range between the current input and the previous input, α represents the adjustment factor, which is used to control the influence of the change rate on the dynamic change of the weight, and ∈ represents a small value, usually 10 -6 or 10 -8 , to avoid the denominator being zero, max is the maximum value of the number of elements in the formula; Introducing Δ t Dynamically adjusted update gate: With' t =from t (1+β Δ t ) Where t represents the time step, z′ t Represents the introduction of Δ t The output of the dynamically adjusted update gate, β is the dynamic adjustment coefficient, which controls the weight amplification effect of input changes on the update gate, z t Represents the output of the original update gate; Introducing Δ t Dynamically adjusted hidden state: Where t represents the time step, h′ t Represents the introduction of Δ t The output of the dynamically adjusted hidden state, h t-1 represents the hidden state of the previous time step, and ⊙ represents element-wise multiplication.

Images