Search

CN-121997756-A - Substation mountain fire risk early warning method based on cellular automata and deep learning

CN121997756ACN 121997756 ACN121997756 ACN 121997756ACN-121997756-A

Abstract

The invention belongs to the technical field of disaster prevention and meteorological disaster prediction of electric power facilities, and relates to a transformer substation mountain fire risk early warning method based on cellular automata and deep learning, comprising the following steps of firstly, accessing and fusing multi-source data in real time, namely, accessing static data and dynamic data, unifying a space coordinate system of all the data and discretizing the data to the same resolution grid to form a cellular matrix of the cellular automata; defining states and attributes for each cell, setting a mountain fire spreading rule based on a physical empirical formula, and carrying out deterministic physical deduction by combining ignition point positions and initial meteorological conditions based on fused static data. The invention has the remarkable advantages of high early warning precision, focusing the early warning object from the 'region' to the 'specific transformer substation', and realizing the targeted early warning.

Inventors

  • WANG WEI
  • CAI LU
  • XU CHUANG
  • DENG JUN
  • TANG CHAO
  • CAO LIANG
  • ZHANG RUILIANG
  • LV GANG
  • TANG HUADONG
  • WANG XIANG
  • TIAN JING
  • XING FANGBO
  • TIAN MAOCHENG
  • CHEN LIANG

Assignees

  • 中国南方电网有限责任公司超高压输电公司贵阳局

Dates

Publication Date
20260508
Application Date
20260128

Claims (8)

  1. 1. The substation mountain fire risk early warning method based on cellular automata and deep learning is characterized by comprising the following steps of: step one, multi-source data real-time access and fusion processing, which is to access static data and dynamic data, unify a space coordinate system of all the data and discretize the data to a grid with the same resolution to form a cellular matrix of a cellular automaton; Defining states and attributes for each cell, setting a mountain fire spreading rule based on a physical empirical formula, and carrying out deterministic physical deduction by combining ignition point positions and initial meteorological conditions based on fused static data; Step three, dynamic parameter real-time correction based on deep learning, namely constructing a lightweight convolutional neural network as a corrector, inputting a real-time dynamic characteristic tensor, outputting a spreading speed correction factor and a wind direction weight correction matrix, and dynamically correcting deduction parameters of the cellular automaton at preset time steps at intervals; And step four, risk assessment and early warning information generation, namely starting a dynamically corrected cellular automaton model to perform multi-round Monte Carlo simulation, generating a dynamic risk level by combining threat probability, expected arrival time and fire scene intensity, and outputting a structured early warning report.
  2. 2. The method for early warning the mountain fire risk of the transformer substation based on the cellular automata and the deep learning, which is disclosed in claim 1, wherein the static data in the first step comprises a high-resolution digital elevation model with the peripheral radius of the transformer substation ranging from 10 kilometers to 20 kilometers, a land utilization/vegetation type graph, a water system distribution graph, a road distribution graph and accurate geographic coordinates and enclosure boundaries of the transformer substation, and the dynamic data comprises local real-time wind speed, wind direction, temperature and humidity data provided by microclimate stations, future 6-72 hours wind field and precipitation probability forecast data provided by regional weather forecast grids, and vegetation indexes, ground surface temperature and hot spot information provided by satellite remote sensing.
  3. 3. The substation mountain fire risk early warning method based on cellular automata and deep learning of claim 1 is characterized in that in the data fusion process of the first step, grid resolution is set to be 30m multiplied by 30m, a WGS-84 space coordinate system is uniformly adopted, and gridding alignment of data from different sources is achieved through a bilinear interpolation method.
  4. 4. The substation mountain fire risk early warning method based on cellular automata and deep learning according to claim 1, wherein the state of the cells in the second step comprises unburnt, burnt/flame retardant, the cell attributes comprise fuel type, fuel humidity, altitude, gradient and slope direction, key parameters in a mountain fire spreading rule comprise spreading speed, spreading direction and ignition probability, the spreading speed is calculated based on Rothermel model basic formulas, the initial value is determined by the fuel model type and gradient, the spreading direction is influenced by the wind direction and the topography, and the ignition probability is determined by comprehensive lightning landing area, satellite hot spots and artificial fire source density historical data.
  5. 5. The substation mountain fire risk early warning method based on cellular automata and deep learning according to claim 1 is characterized in that in the third step, the input characteristic tensor of a convolutional neural network is 5km multiplied by 5km grid data centering on a substation, the input characteristic tensor comprises 8 channels of real-time wind speed, wind direction, vegetation humidity and adjacent cellular CA states of each cellular, an propagation speed correction factor alpha and a wind direction weight correction matrix W of each cellular are output, space-time evolution sequence data of historical mountain fire cases are adopted in model training, jaccard coefficients simulating a fire scene boundary and a real fire scene boundary are used as loss functions, and supervision training is carried out through an Adam optimizer.
  6. 6. The substation mountain fire risk early warning method based on cellular automata and deep learning according to claim 1, wherein the dynamic correction period in the third step is consistent with the deduction time step of the cellular automata, and is set to 15 minutes, namely after each time step of physical deduction is completed, correction parameters are updated based on latest real-time data for the next round of deduction.
  7. 7. The substation mountain fire risk early warning method based on cellular automata and deep learning, which is disclosed in claim 1, is characterized in that the Monte Carlo simulation number of the fourth step is not less than 500, weather forecast uncertainty and random disturbance of fuel humidity are introduced in the simulation process, the disturbance range comprises wind speed + -10%, wind direction + -15 degrees and humidity + -5 degrees, threat probability is the proportion of the number of times that any point of a substation enclosure is covered by a probability fire field to the total simulation number, and a probability threshold is set to be 30%.
  8. 8. The substation mountain fire risk early warning method based on cellular automata and deep learning, which is characterized in that risk class in the fourth step is divided into four classes of low, medium, high and extremely high, threat probability P, expected arrival time T and fire intensity I are comprehensively considered, wherein the fire intensity is estimated by the product of fuel load and spreading speed, and the structured early warning report comprises risk class, threat probability of each future time point, expected mountain fire arrival time range, main threat direction and visualized risk thermodynamic diagram.

Description

Substation mountain fire risk early warning method based on cellular automata and deep learning Technical Field The invention belongs to the technical field of disaster prevention and meteorological disaster prediction of electric power facilities, and particularly relates to a substation forest fire risk early warning method based on cellular automata and deep learning. Background With the continuous increase of power demand, the construction of the transformer substation gradually expands to suburbs and mountain forest lands, and the risk of transformer substation outage and even equipment damage caused by external mountain fires is increasingly prominent, so that the safe and stable operation of a power system is seriously threatened. The existing mountain fire early warning technology of the transformer substation mainly depends on two types of modes, namely, macroscopic weather early warning based on administrative division or grid (such as 10km multiplied by 10 km) is low in spatial resolution, small-scale topography and vegetation difference around the transformer substation cannot be embodied, early warning positions are wide, risks of specific transformer substations are difficult to accurately locate, alarming realized through observation whistle or satellite hot spot monitoring belongs to 'post' early warning, alarming is triggered after open fire is found, early warning time is usually only tens of minutes, and window period reserved for emergency preparation of the transformer substation is seriously insufficient. The prior art has the following core defects: The early warning precision is insufficient, the probability and time of the mountain fire spreading to a specific transformer substation cannot be quantized, so that an emergency response decision lacks scientific basis, and the problems of 'wolf coming' effect or untimely response and out-of-place problem easily occur; the timeliness is poor, the early warning based on hot spot monitoring is triggered after the fire is formed, the early warning advance is insufficient, and the time requirement of emergency treatment of the transformer substation is difficult to meet; The dynamic adaptability is weak, the traditional fire risk model is static or semi-static, and key meteorological data such as real-time local wind speed and direction mutation, humidity change and the like cannot be effectively fused, so that the deviation between fire deduction and actual conditions is larger. The Cellular Automaton (CA) is used as a classical physical mechanism model, can effectively simulate the space-time evolution process of a complex system, is suitable for physical deduction of mountain fire spreading, and has strong data fitting and real-time correction capability, and model parameters can be dynamically adjusted to adapt to environmental changes. However, how to organically combine the two, and realizing accurate dynamic deduction of mountain fire spreading and substation targeting early warning, is still a technical problem to be solved currently. The invention aims to overcome the defects of the prior art, provides a substation mountain fire risk early warning method based on cellular automata and deep learning, realizes the crossing from 'regional early warning' to 'site targeting early warning', and provides accurate decision support for substation grading emergency response. Disclosure of Invention The invention aims to provide a bank project early warning method of a knowledge graph, which solves the problems existing in the prior art. The aim of the invention can be achieved by the following technical scheme: the substation mountain fire risk early warning method based on cellular automata and deep learning comprises the following steps of firstly, accessing and fusing multi-source data in real time, namely accessing static data and dynamic data, unifying a space coordinate system of all the data and discretizing the data to a grid with the same resolution to form a cellular matrix of the cellular automata; Defining states and attributes for each cell, setting a mountain fire spreading rule based on a physical empirical formula, and carrying out deterministic physical deduction by combining ignition point positions and initial meteorological conditions based on fused static data; Step three, dynamic parameter real-time correction based on deep learning, namely constructing a lightweight convolutional neural network as a corrector, inputting a real-time dynamic characteristic tensor, outputting a spreading speed correction factor and a wind direction weight correction matrix, and dynamically correcting deduction parameters of the cellular automaton at preset time steps at intervals; And step four, risk assessment and early warning information generation, namely starting a dynamically corrected cellular automaton model to perform multi-round Monte Carlo simulation, generating a dynamic risk level by combining threat probability, expected arri