Transformer substation forest fire risk early warning method based on cellular automaton and deep learning

Abstract

The invention belongs to the technical field of electric power facility disaster prevention and meteorological disaster prediction, and relates to a substation forest fire risk early warning method based on a cellular automaton and deep learning, which comprises the following steps: step 1, multi-source data real-time access and fusion processing: accessing static data and dynamic data, and carrying out multi-source data fusion processing; unifying all data into a space coordinate system and discretizing the data into grids with the same resolution ratio to form a cellular matrix of the cellular automaton; and 2, forest fire spreading physical modeling based on a cellular automaton: defining a state and an attribute for each cell, setting a forest fire spreading rule based on a physical empirical formula, and carrying out deterministic physical deduction by taking fused static data as a basis and combining an ignition point position and an initial meteorological condition. The method has the following remarkable advantages that the early warning precision is high, an early warning object is focused to a specific transformer substation from a region, and targeted early warning is achieved;

Description

Technical Field

[0001] This invention belongs to the field of power facility disaster prevention and meteorological disaster prediction technology, specifically, it relates to a substation wildfire risk early warning method based on cellular automata and deep learning. Background Technology

[0002] With the continuous growth of electricity demand, substation construction is gradually expanding to suburban and mountainous areas. The risk of substation shutdowns and even equipment damage caused by wildfires is becoming increasingly prominent, seriously threatening the safe and stable operation of the power system. Existing substation wildfire early warning technologies mainly rely on two types of methods: one is macro-meteorological warnings issued based on administrative divisions or grids (such as 10km×10km). This type of warning has low spatial resolution and cannot reflect small-scale differences in terrain and vegetation around the substation. The warning location is broad and it is difficult to accurately locate the risk of specific substations. The other is alarms achieved through lookout posts or satellite hotspot monitoring. This is a "post-event" warning, which is triggered only after an open fire is discovered. The warning time is usually only tens of minutes, leaving insufficient window of opportunity for substation emergency preparation.

[0003] The existing technology has the following core defects: Insufficient early warning accuracy: It is impossible to quantify the probability and time of wildfires spreading to a specific substation, resulting in a lack of scientific basis for emergency response decisions and a tendency to have a "crying wolf" effect or untimely and inadequate response. Poor timeliness: Early warnings based on hotspot monitoring are only triggered after the fire has developed, resulting in insufficient lead time and making it difficult to meet the time requirements for emergency response at substations; Weak dynamic adaptability: Traditional fire risk models are mostly static or semi-static, which cannot effectively integrate key meteorological data such as real-time local wind speed and direction changes and humidity changes, resulting in a large deviation between fire simulation and actual situation.

[0004] Cellular automata (CA), as a classic physical mechanism model, can effectively simulate the spatiotemporal evolution of complex systems and is suitable for physical extrapolation of wildfire spread. Deep learning models, on the other hand, possess powerful data fitting and real-time correction capabilities, and can dynamically adjust model parameters to adapt to environmental changes. However, how to organically integrate the two to achieve accurate dynamic extrapolation of wildfire spread and targeted early warning of substations remains a pressing technical challenge.

[0005] This invention aims to overcome the shortcomings of existing technologies and provide a substation wildfire risk early warning method based on cellular automata and deep learning, realizing the leap from "regional early warning" to "site-targeted early warning" and providing accurate decision support for substation graded emergency response. Summary of the Invention

[0006] The purpose of this invention is to provide a knowledge graph-based early warning method for bank projects, which solves the problems existing in the prior art.

[0007] The objective of this invention can be achieved through the following technical solutions: The substation wildfire risk early warning method based on cellular automata and deep learning includes the following steps: Step 1, real-time access and fusion processing of multi-source data: access static data and dynamic data, unify all data into a spatial coordinate system and discretize them to the same resolution grid to form the cell matrix of cellular automata. Step 2: Physical modeling of wildfire spread based on cellular automata: Define the state and attributes for each cell, set wildfire spread rules based on physical empirical formulas, and perform deterministic physical deduction based on the fused static data and the location of the ignition point and the initial meteorological conditions. Step 3: Real-time dynamic parameter correction based on deep learning: Construct a lightweight convolutional neural network as a corrector, input a real-time dynamic feature tensor, and output a propagation speed correction factor and a wind direction weight correction matrix. The inference parameters of the cellular automata are dynamically corrected at preset time steps. Step 4: Risk Assessment and Early Warning Information Generation: Activate the dynamically corrected cellular automata model to conduct multiple rounds of Monte Carlo simulations, combine threat probability, estimated arrival time, and fire intensity to generate dynamic risk levels, and output a structured early warning report.

[0008] Preferably, the static data in step one includes a high-resolution digital elevation model of the substation with a radius of 10-20 kilometers, land use / vegetation type map, water system distribution map, road distribution map, and the precise geographic coordinates and perimeter wall boundary of the substation; the dynamic data includes local real-time wind speed, wind direction, temperature, and humidity data provided by micro-weather stations, wind field and precipitation probability forecast data for the next 6-72 hours provided by regional meteorological forecast grids, and vegetation index, surface temperature, and hotspot information provided by satellite remote sensing.

[0009] Preferably, in the data fusion process of step one, the grid resolution is set to 30m×30m, the WGS-84 spatial coordinate system is uniformly adopted, and the raster alignment of data from different sources is achieved by bilinear interpolation.

[0010] Preferably, in step two, the cell states include unburned, burning, and ignited / flame-retardant, and the cell attributes include fuel type, fuel moisture, altitude, slope, and aspect. The key parameters in the wildfire spread rules include spread rate, spread direction, and ignition probability. The spread rate is calculated based on the basic formula of the Rothermel model, and the initial value is determined by the fuel model type and slope. The spread direction is affected by the prevailing wind direction and terrain. The ignition probability is determined by combining historical data on lightning strike areas, satellite hotspots, and anthropogenic fire source density.

[0011] Preferably, in step three, the input feature tensor of the convolutional neural network is a 5km×5km grid data centered on the substation, containing eight channels for each cell: real-time wind speed, wind direction, vegetation humidity, and CA state of neighboring cells; the output is a spread rate correction factor α and a wind direction weight correction matrix W for each cell; the model training uses spatiotemporal evolution sequence data of historical wildfire cases, with the Jaccard coefficient between the simulated fire boundary and the real fire boundary as the loss function, and supervised training is performed through the Adam optimizer.

[0012] Preferably, the dynamic correction period in step three is consistent with the deduction time step of the cellular automaton, and is set to 15 minutes. That is, after each time step of physical deduction is completed, the correction parameters are updated based on the latest real-time data for the next round of deduction.

[0013] Preferably, the number of Monte Carlo simulations in step four is no less than 500 rounds. During the simulation, weather forecast uncertainty and random disturbances of fuel humidity are introduced. The disturbance range includes wind speed ±10%, wind direction ±15°, and humidity ±5%. The threat probability is the proportion of the number of times any point on the substation perimeter wall is covered by a probabilistic fire field to the total number of simulations, and the probability threshold is set at 30%.

[0014] Preferably, the risk level in step four is divided into four levels: low, medium, high, and extremely high, taking into account the threat probability P, the estimated arrival time T, and the fire intensity I. The fire intensity is estimated by the product of fuel load and spread rate. The structured early warning report includes the risk level, the threat probability at each future time point, the estimated wildfire arrival time range, the main threat direction, and a visualized risk heat map.

[0015] The beneficial effects of the present invention: The present invention has the following significant advantages: High early warning accuracy: It focuses the early warning target from "area" to "specific substation" to achieve targeted early warning; it integrates 30m×30m high-resolution geographic data and local real-time meteorological data to improve the spatial resolution of the assessment from the kilometer level to the hundred-meter level, which can accurately capture small-scale environmental differences around the substation. With a large lead time for early warning: Based on the dynamic physical simulation of wildfire spread, a probabilistic and time-lined early warning can be given when the fire point is several kilometers away from the substation, giving the substation several hours of valuable preparation time to activate its emergency plan; Strong dynamic adaptive capability: It adopts a hybrid architecture of "cellular automata + deep learning", which not only ensures the physical rationality of wildfire spread projection through CA model, but also corrects parameters in real time through deep learning model, effectively responding to key environmental changes such as sudden changes in wind speed and wind direction, overcoming the shortcomings of fixed parameters in pure physical model and poor interpretability in pure data-driven model. The system provides rich decision support information: it not only provides qualitative assessments of risk levels, but also outputs multi-dimensional quantitative information such as threat probability, estimated arrival time, and main threat direction, supporting maintenance personnel to make precise emergency decisions in a graded and phased manner, thereby improving the pertinence and effectiveness of substation disaster prevention and mitigation. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a system block diagram of the substation wildfire risk early warning method based on cellular automata and deep learning according to the present invention. Figure 2 This is a schematic diagram of the cellular structure and state transition rules of the CA model in the substation wildfire risk early warning method based on cellular automata and deep learning of this invention.

[0018] Figure 3 This is a wildfire spread probability cloud map and a substation risk diagram generated by Monte Carlo simulation in the substation wildfire risk early warning method based on cellular automata and deep learning of this invention.

[0019] Figure 4 This is a schematic diagram illustrating a visual output example of the early warning report in the substation wildfire risk early warning method based on cellular automata and deep learning of this invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figures 1-4 As shown, the substation wildfire risk early warning method based on cellular automata and deep learning of the present invention includes the following steps: Step 1: Real-time Access and Fusion Processing of Multi-Source Data: This step involves accessing both static and dynamic data, unifying all data into a single spatial coordinate system, and discretizing them onto a grid of the same resolution to form a cellular matrix of cellular automata. The static data in Step 1 includes a high-resolution digital elevation model of the substation's perimeter (within a radius of 10-20 km), land use / vegetation type maps, water system distribution maps, road distribution maps, and the substation's precise geographic coordinates and perimeter boundary. The dynamic data includes real-time local wind speed, wind direction, temperature, and humidity data provided by micro-meteorological stations; wind field and precipitation probability forecasts for the next 6-72 hours provided by regional meteorological forecast grids; and vegetation indices, surface temperature, and hotspot information provided by satellite remote sensing. During the data fusion process in Step 1, the grid resolution is set to 30m × 30m, and the WGS-84 spatial coordinate system is uniformly used. Raster alignment of data from different sources is achieved through bilinear interpolation.

[0022] Step 2: Physical Modeling of Wildfire Spread Based on Cellular Automata: Define the state and attributes for each cell, set wildfire spread rules based on empirical physical formulas, and perform deterministic physical deduction based on the fused static data and the location of ignition points and initial meteorological conditions. The cell states in Step 2 include unburned, burning, and ignited / flame-retardant, and the cell attributes include fuel type, fuel moisture, altitude, slope, and aspect. The key parameters in the wildfire spread rules include spread rate, spread direction, and ignition probability. The spread rate is calculated based on the basic formula of the Rothermel model, and the initial value is determined by the fuel model type and slope. The spread direction is affected by the prevailing wind direction and terrain. The ignition probability is determined by combining historical data on lightning strike areas, satellite hotspots, and anthropogenic fire source density.

[0023] Step 3: Real-time dynamic parameter correction based on deep learning: A lightweight convolutional neural network is constructed as the corrector. It takes a real-time dynamic feature tensor as input and outputs a spread rate correction factor and a wind direction weight correction matrix. The inference parameters of the cellular automata are dynamically corrected at preset time steps. In Step 3, the input feature tensor of the convolutional neural network is a 5km×5km grid of data centered on the substation, containing eight channels for each cell: real-time wind speed, wind direction, vegetation humidity, and CA state of neighboring cells. The output is a spread rate correction factor α and a wind direction weight correction matrix W for each cell. Model training uses spatiotemporal evolution sequence data from historical wildfire cases, with the Jaccard coefficient between the simulated fire boundary and the real fire boundary as the loss function, and supervised training using the Adam optimizer. The dynamic correction cycle in Step 3 is consistent with the inference time step of the cellular automata, set to 15 minutes. That is, after each time step of physical inference, the correction parameters are updated based on the latest real-time data for the next round of inference.

[0024] Step 4: Risk Assessment and Early Warning Information Generation: A dynamically corrected cellular automata model is initiated for multiple rounds of Monte Carlo simulation. Dynamic risk levels are generated by combining threat probability, estimated arrival time, and fire intensity, and a structured early warning report is output. The Monte Carlo simulation in Step 4 involves no fewer than 500 rounds. Uncertainty in weather forecasts and random disturbances in fuel humidity are introduced during the simulation. These disturbances include wind speed ±10%, wind direction ±15°, and humidity ±5%. The threat probability is the proportion of times any point on the substation perimeter wall is covered by a probabilistic fire out of the total number of simulations, with a probability threshold of 30%. The risk levels in Step 4 are divided into four levels: low, medium, high, and extremely high, comprehensively considering the threat probability P, estimated arrival time T, and fire intensity I. Fire intensity is estimated by multiplying fuel load and spread rate. The structured early warning report includes the risk level, threat probability at various future time points, estimated wildfire arrival time range, main threat direction, and a visualized risk heat map.

[0025] The specific implementation method is as follows: Example scenario setting: Target station: A 500kV mountain substation in southern China (virtual name "Yuntai Station"); Coordinates: 112.738°E, 25.031°N, altitude 420m; Boundary: approximately 280m wide east-west, 320m deep north-south, with a 10m firebreak outside the surrounding walls; Fire source: At 14:30 on March 18, 2025, satellite VIIRS detected a hotspot (112.66°E, 25.08°N) 8.2km northwest of Yuntai Station; Meteorology: At 14:30, the micro-weather station measured a westerly wind of 6.2m / s, relative humidity of 28%, and temperature of 22℃; WRF forecasts continued westerly winds for the next 6 hours, with a wind speed of 5-7m / s; Vegetation: The western 0-4km is mainly composed of Masson pine (fuel model 4), NDVI 0.62, LST 301K.

[0026] Step 1: Data Preparation and Fusion. Static data includes: 30m resolution ASTERGDEM (slope 5-25°, west-facing), 10m resolution Southern Power Grid Remote Sensing Center 2024 land use map (pine forest 46%, shrubs 26%, grassland 18%), OpenStreetMap 2024Q4 road / water system data (rasterized after 30m buffer), and precise coordinates of substations and perimeter boundary vector data; Dynamic data includes: 1-minute sampling data from 3 micro-weather stations (TCP pushed to the edge server), WRF-ARW 3km resolution forecast data updated every 1 hour, MODIS / Terra daily NDVI and LST data, and VIIRS 375m resolution active fire point data updated every 15 minutes; Data fusion: The WGS-84 coordinate system is uniformly adopted, and all data are rasterized into 30m×30m grids using bilinear interpolation, forming a 15km×15km (501×501 cell) cell matrix.

[0027] Step 2: Cellular automata model initialization. Cell state encoding: 0 = unburned, 1 = burning, 2 = burned / flame-retardant; Fuel model: 13 types of Albini-Scott model are used, with a local correction coefficient of 0.92; Initial parameters: Initial spread velocity R0 is calculated using the Rothermel formula, averaging 0.38 m / s, with a slope factor of 1.18 and a default wind direction weight matrix of 0.65; Ignition point setting: The hotspot location discovered by satellite VIIRS is set as the initial ignition point, with the state marked as 1 (burning).

[0028] Step 3: Corrector Training and Deployment Convolutional Neural Network Structure: Input layer 64×64×8 channels, intermediate layers are 64×64×16, 64×64×32, and 64×64×16 convolutional layers respectively, output layer 64×64×2 (α, W); Model Training: Using 184 time-phase data from 3 real wildfires from 2019 to 2023, Jaccard loss function value 0.186, deployed on NVIDIA Jetson AGX Orin edge server, inference time 0.8s / inference; Dynamic Correction: The correction parameters are updated every 15 minutes based on the latest real-time data, adjusting the spread speed and wind direction weights of the CA model.

[0029] Step 4: Real-time Early Warning Simulation and Result Output Monte Carlo Simulation: 500 simulations were conducted, incorporating random disturbances of wind speed ±10%, wind direction ±15°, and humidity ±5%, with a time step of 15 minutes; Simulation results: At 15:00 (0.5h after the fire), a 6-hour probability cloud map was generated, with a 3-hour threat probability of 52%, the earliest arrival time being 2h20min, the latest arrival time being 4h10min, and the western perimeter wall having the highest risk; Early warning output: At 15:02, a structured early warning report with a "high" risk level was generated, including a risk heat map, timeline, and detailed parameters, which was pushed to the provincial dispatch center via JSON message and displayed on the visualization platform.

[0030] In the description of this specification, the references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0031] The above description is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in the claims, they should all fall within the protection scope of the present invention.

Claims

1. A substation wildfire risk early warning method based on cellular automata and deep learning, characterized by: Includes the following steps: Step 1: Real-time access and fusion processing of multi-source data: Access static and dynamic data, unify all data into a spatial coordinate system and discretize them to the same resolution grid to form the cell matrix of cellular automata; Step 2: Physical modeling of wildfire spread based on cellular automata: Define the state and attributes for each cell, set wildfire spread rules based on physical empirical formulas, and perform deterministic physical deduction based on the fused static data and the location of the ignition point and the initial meteorological conditions. Step 3: Real-time dynamic parameter correction based on deep learning: Construct a lightweight convolutional neural network as a corrector, input a real-time dynamic feature tensor, and output a propagation speed correction factor and a wind direction weight correction matrix. The inference parameters of the cellular automata are dynamically corrected at preset time steps. Step 4: Risk Assessment and Early Warning Information Generation: Activate the dynamically corrected cellular automata model to conduct multiple rounds of Monte Carlo simulations, combine threat probability, estimated arrival time, and fire intensity to generate dynamic risk levels, and output a structured early warning report.

2. The substation wildfire risk early warning method based on cellular automata and deep learning according to claim 1, wherein the static data in step one includes a high-resolution digital elevation model of the substation with a radius of 10-20 kilometers, land use / vegetation type map, water system distribution map, road distribution map, and the precise geographic coordinates and perimeter wall boundary of the substation; the dynamic data includes local real-time wind speed, wind direction, temperature, and humidity data provided by micro-meteorological stations, wind field and precipitation probability forecast data for the next 6-72 hours provided by regional meteorological forecast grids, and vegetation index, surface temperature, and hotspot information provided by satellite remote sensing.

3. The substation wildfire risk early warning method based on cellular automata and deep learning according to claim 1, characterized in that: In the data fusion process of step one, the grid resolution is set to 30m×30m, the WGS-84 spatial coordinate system is used uniformly, and the raster alignment of data from different sources is achieved by bilinear interpolation.

4. The substation wildfire risk early warning method based on cellular automata and deep learning according to claim 1, characterized in that: In step two, the cell states include unburned, burning, and ignited / flame-retardant, and the cell attributes include fuel type, fuel moisture, altitude, slope, and aspect. The key parameters in the wildfire spread rules include spread rate, spread direction, and ignition probability. The spread rate is calculated based on the basic formula of the Rothermel model, and the initial value is determined by the fuel model type and slope. The spread direction is affected by the prevailing wind direction and terrain. The ignition probability is determined by combining historical data on lightning strike areas, satellite hotspots, and anthropogenic fire source density.

5. The substation wildfire risk early warning method based on cellular automata and deep learning according to claim 1, characterized in that: In step three, the input feature tensor of the convolutional neural network is a 5km×5km grid data centered on the substation, containing eight channels for each cell: real-time wind speed, wind direction, vegetation humidity, and CA state of neighboring cells. The output is a spread rate correction factor α and a wind direction weight correction matrix W for each cell. The model training uses spatiotemporal evolution sequence data of historical wildfire cases, with the Jaccard coefficient between the simulated fire boundary and the real fire boundary as the loss function, and supervised training is performed through the Adam optimizer.

6. The substation wildfire risk early warning method based on cellular automata and deep learning according to claim 1, characterized in that: The dynamic correction cycle in step three is consistent with the deduction time step of the cellular automata, which is set to 15 minutes. That is, after each time step of physical deduction is completed, the correction parameters are updated based on the latest real-time data for the next round of deduction.

7. The substation wildfire risk early warning method based on cellular automata and deep learning according to claim 1, characterized in that: The Monte Carlo simulation in step four is no less than 500 rounds. During the simulation, weather forecast uncertainty and random disturbance of fuel humidity are introduced. The disturbance range includes wind speed ±10%, wind direction ±15°, and humidity ±5%. The threat probability is the proportion of the number of times any point on the substation wall is covered by a probabilistic fire field out of the total number of simulations. The probability threshold is set at 30%.

8. The substation wildfire risk early warning method based on cellular automata and deep learning according to claim 1, characterized in that: The risk level in step four is divided into four levels: low, medium, high, and extremely high, taking into account the threat probability P, the estimated arrival time T, and the fire intensity I. The fire intensity is estimated by the product of fuel load and spread rate. The structured early warning report includes the risk level, the threat probability at each future time point, the estimated wildfire arrival time range, the main threat direction, and a visualized risk heat map.

Images