Forest fire spreading intelligent prediction and active intervention control method and system

Abstract

The invention provides a forest fire spreading intelligent prediction and active intervention control method and system, and relates to the technical field of forest fire management. The intelligent prediction and active intervention system for forest fire spreading comprises a data acquisition module, a data preprocessing module, an intelligent prediction module, a model optimization and verification module, an active intervention decision module, an intervention strategy optimization module, a real-time monitoring and evaluation module, a user interaction interface module and a communication module. Through a series of core technologies of highly integrating multi-dimensional data acquisition, intelligent data analysis and prediction, automatic intervention control, dynamic optimization adjustment and the like, efficient and accurate prediction of the forest fire spreading trend is realized. The system not only can comprehensively collect multi-dimensional data including meteorological conditions, landforms, vegetation distribution and the like in real time, but also utilizes advanced deep learning and machine learning algorithms to intelligently analyze the data, so that a high-precision forest fire spreading prediction model is constructed.

Description

Technical Field

[0001] The present invention relates to the technical field of forest fire management, and in particular to a method and system for intelligent prediction and active intervention control of forest fire spread. Background Art

[0002] In the field of forest fire management, traditional methods of forest fire spread prediction and prevention and control face many challenges. On the one hand, because forest fire spread is affected by a variety of complex factors such as meteorological conditions, topography, and vegetation types, traditional prediction methods based on experience or a single data source are difficult to accurately capture the interaction between these factors, resulting in greater uncertainty in the prediction results. On the other hand, after a fire occurs, traditional manual intervention and response mechanisms often lag behind, making it difficult to effectively control the spread of the fire in a timely manner, thus posing a serious threat to forest resources, the ecological environment, and even the safety of people's lives and property.

[0003] With the rapid development of technologies such as artificial intelligence and big data, new solutions have been provided for forest fire management. By deeply mining the intrinsic connections between multidimensional data (including meteorology, topography, vegetation, historical forest fire records, human activities, etc.), and using advanced algorithms such as deep reinforcement learning to build intelligent prediction models, the accuracy and timeliness of forest fire spread predictions can be significantly improved. At the same time, combined with modern technical means such as intelligent Internet of Things and drone inspections, all-weather and all-round monitoring of forest areas can be achieved, providing strong support for timely and effective active intervention and control.

[0004] Although there are some forest fire management systems on the market, most of these systems focus on a single monitoring or early warning function, and lack the organic combination of intelligent prediction and active intervention control. To this end, the present invention proposes an innovative method and system for intelligent prediction and active intervention control of forest fire spread. The system aims to achieve full-chain management of forest fire spread by deeply integrating advanced technology and multi-dimensional data. It not only covers prediction and intervention, but also includes evaluation, forming a closed-loop intelligent management system to minimize the losses caused by forest fires. Summary of the invention

[0005] 1. Technical issues to be solved

[0006] In view of the shortcomings of the existing technology, the present invention provides a method and system for intelligent prediction and active intervention control of forest fire spread, which solves the problem that traditional manual intervention and response mechanisms often have lags and are difficult to control the spread of fire in a timely and effective manner, thus posing a serious threat to forest resources, the ecological environment and even the safety of people's lives and property.

[0007] (II) Technical solution

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: a forest fire spread intelligent prediction and active intervention system, including: a data acquisition module, a data preprocessing module, an intelligent prediction module, a model optimization and verification module, an active intervention decision module, an intervention strategy optimization module, a real-time monitoring and evaluation module, a user interaction interface module and a communication module.

[0009] A control method for a forest fire spread intelligent prediction and active intervention system comprises the following steps:

[0010] S1. Meteorological data acquisition: real-time and historical meteorological data including temperature, humidity, wind speed, wind direction and precipitation are acquired through meteorological stations and satellite remote sensing. Topographic and geomorphological acquisition: topographic maps and DEM digital elevation model data of the region are acquired using GIS system to analyze the characteristics of altitude, slope, aspect and soil type. Vegetation distribution: vegetation type, density, moisture content and flammability data are acquired by combining satellite remote sensing images and ground surveys. Historical forest fire records: detailed records of historical fires are acquired from forestry departments or relevant institutions, including the location of the fire point, the path of spread, the intensity of the fire and the time of extinguishing. Human activities: population density, road distribution and fire source management records are evaluated through population census and transportation network data.

[0011] S2. Data collection and preprocessing,For data with a distribution close to normal, the standard deviation range is used for outlier detection. The calculation process is as follows:

[0012] To calculate the mean and standard deviation, first calculate the mean μ and standard deviation σ of the data set:

[0013]

[0014]

[0015] Where N is the number of values ​​in the data set, x i is the i-th value in the data set;

[0016] Noise removal, missing value filling and standardization processing. First, the input layer receives the preprocessed raw data sequence, which can be a time series signal, such as an audio signal or sensor data; next, the data passes through the first convolution layer, which uses 32 3x3 convolution kernels to extract local spatial features, and then the data is reduced in dimension through the maximum pooling layer to reduce the redundant information of the features; the second convolution layer uses 64 3x3 convolution kernels to further extract higher-level spatial features, and the pooling layer is also used to reduce the feature dimension; the feature maps processed by the convolution layer and the pooling layer are flattened and input to the LSTM layer. The LSTM layer consists of two stacked LSTM layers, each containing 128 units, responsible for capturing time Dynamic features in the sequence; Subsequently, these features are further processed by the fully connected layer to generate the \(k\)th noise pattern \(P_k(t)\) and the corresponding weight coefficient \(\alpha_k\). Finally, the output layer applies these noise patterns and weight coefficients to calculate the denoised data values; During the training process, we use a standardized historical data set as the training set and adopt the mean square error MSE as the loss function. The model parameters are adjusted through the back propagation algorithm to minimize the error between the denoised data and the actual data; This deep learning model that combines CNN and LSTM can effectively and adaptively identify and remove noise in the data, thereby significantly improving the quality and reliability of the data;

[0017] The noise removal formula is:

[0018]

[0019] in, is the data value after removing noise at time point t; X t is the value of the original data at time point t; M i (t) is the value of the i-th noise pattern at time point t, generated by the deep learning model; W i is the corresponding weight coefficient, which is learned through the training process; N is the number of noise patterns;

[0020] Noise Mode M i The calculation of (t) depends on the output of the deep learning model. Assuming that the noise pattern is generated by the features predicted by the model, it can be expressed as:

[0021] M i (t) = f i (CNN(X t )·RNN)

[0022] Among them, CNN (X t ) is the original data X t The features extracted by the convolutional layer are applied; RNN is the time series features processed by the LSTM layer; fi (·) is the i-th noise pattern generated by the fully connected layer or other output layers;

[0023] Missing value filling adopts a new spatiotemporal interpolation method, combines time series analysis and spatial interpolation technology, integrates time and space features to predict missing values, and builds a spatiotemporal model to predict and fill missing data points. The specific filling formula is:

[0024] FilledValue t·j =f(TimeScriesComponent t , SpatialComponent j )

[0025] Among them, FilledValue t·j represents the filling value at time point t and spatial position j, and f is the spatiotemporal interpolation function, which comprehensively considers the temporal continuity and spatial correlation of the data; TimeScriesComponent t Represents the value of the time series component at time point t, obtained through time series analysis: SpatialComponent j Represents the value of the spatial component at spatial position j, obtained by spatial interpolation technology;

[0026] The calculation formula for meteorological data is:

[0027]

[0028] in, is the fill value, temperature, humidity at time point t and spatial position s; T t is the time series component, representing the meteorological data and temperature at time point t; H s is the spatial component, representing the meteorological data and humidity at the spatial location s; W s are other spatial features, such as wind speed and wind direction; α T , α H , α W , β and δ are model parameters;

[0029] The calculation formula for topography is:

[0030]

[0031] in, is the filled value, soil moisture, at time point t and spatial location s; E S is the spatial component, representing the altitude data of the spatial position s; S s is the spatial component, indicating the slope of the spatial position s; A sis the spatial component, indicating the slope direction of the spatial position s; α E , α s , α A , β and δ are model parameters;

[0032] The calculation formula for vegetation distribution is:

[0033]

[0034] in, is the filling value at time point t and spatial position s, the vegetation density; V s is the spatial component, indicating the vegetation type and density at the spatial location s; D s is the other spatial component, water content; α v , α D , β and δ are model parameters;

[0035] The calculation formula for historical forest fire records is:

[0036]

[0037] in, is the fill value at time point t and spatial location s, fire intensity; F s is the spatial component, representing the historical fire record and fire intensity at the spatial location s; P s is the other spatial component, the location of the fire point; α F , α P , β and δ are model parameters;

[0038] The calculation formula for human activities is:

[0039]

[0040] in, is the fill value at time point t and spatial location s, population density; P s is the spatial component, representing the population density at spatial location s; R s is the other spatial component, road distribution; α p , α R , β and δ are model parameters;

[0041] According to the actual distribution and characteristics of the data, the normalization parameters are adjusted dynamically. The specific normalization formula is:

[0042]

[0043] Among them, StandardizedData t represents the standardized data at time point t, μ t and σ′ tRespectively represent the mean of the data at time point t and the dynamically adjusted standard deviation, adaptive adjustment strategy:

[0044] Threshold setting: Set the threshold for data distribution changes. When the changes in the real-time mean and standard deviation exceed the set threshold, dynamic adjustment is triggered;

[0045] Adjustment factor: Use an adaptive adjustment factor to update the normalized parameters, which can be calculated based on the changes in historical data;

[0046] The dynamic adjustment formula is:

[0047] μ′ t =μ t-1 +β·(μ t -μ t-1 )

[0048] σ′ t =σ t-1 +β·(σ t -σ t-1 )

[0049] Among them, μ′ t and σ′ t is the adjusted mean and standard deviation, β is the adjustment factor, usually between 0 and 1, which determines the weight of historical and current data on the adjustment. In the data preprocessing stage, the original data is first standardized to ensure that all data have the same dimension and range, thereby improving the stability and efficiency of model training. Then, the unnecessary noise in the data is removed through the denoising process to ensure that the model can focus on useful signals. Then, the processed data is divided into a training set and a test set, where the training set is used for model training and the test set is used to evaluate the model performance. The following feature extraction process includes multiple steps:

[0050] First, the data is passed through the first convolutional layer to extract spatial features, which captures the basic local information in the data;

[0051] Then, the data enters the second convolutional layer, which further extracts higher-dimensional spatial features to obtain more complex feature representations. The feature map output by the convolutional layer is processed by the maximum pooling layer. The pooling operation reduces the feature dimension and the computational burden by selecting the most significant features, while retaining important spatial information. Next, the feature map after the pooling layer is flattened and sent to the fully connected layer for further processing and high-level feature fusion; the fully connected layer comprehensively processes the flattened feature map to generate feature representations for subsequent analysis; then, these feature representations are input to the LSTM layer, which processes the output of the fully connected layer, captures dynamic changes in the time series, and identifies the temporal characteristics and patterns of the data; in this way, the model can comprehensively consider the characteristics of both space and time for in-depth analysis and prediction;

[0052] In terms of innovation and optimization of the prediction model, the input layer receives carefully pre-processed multi-dimensional data, which are first subjected to spatial feature extraction through multiple convolutional layers. The convolution operation captures multi-scale spatial features in the data through convolution kernels of different sizes. Next, the pooling layer is used to reduce the dimension and extract the most significant features, reduce the amount of calculation and prevent overfitting. After the convolution and pooling operations, the feature map is flattened and sent to one or more fully connected layers, which are used to further fuse the spatial features and perform high-level abstraction. Next, these features are subjected to time series analysis through the LSTM layer to capture the temporal relationship and dynamic changes in the data.

[0053] Input layer: receives preprocessed multidimensional data, such as time series signals or image data;

[0054] Convolutional layer, CNN part

[0055] First convolutional layer: convolution kernel: size is 3×3, number of layers: 3 convolutional layers;

[0056] formula:

[0057]

[0058] Among them, C1(x,y) is the feature map after the convolution operation, W i,j is the convolution kernel weight, X x+i,y+j is the input data value, b is the bias term;

[0059] Second convolutional layer: convolution kernel: size 5×5,

[0060] formula:

[0061]

[0062] Where C2(x,y) is the feature map after the convolution operation;

[0063] Pooling layer:

[0064] Type: max pooling, pooling size 2×2;

[0065] formula:

[0066]

[0067] Among them, P i,j It is the feature map after pooling;

[0068] Fully connected layers, number of layers: one or more fully connected layers;

[0069] Function: to fuse and abstract the features extracted by the convolution kernel pooling layer at a high level;

[0070] formula:

[0071]

[0072] Among them, F k is the output of the fully connected layer, A i is the activation value of the previous layer, W ik is the weight of the fully connected layer, b k is the bias term;

[0073] LSTM layer, RNN part, number of layers: 2 LSTM layers; number of units: 50 LSTM units per layer;

[0074] formula:

[0075] Forget Gate:

[0076] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0077] Input Gate:

[0078] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0079] Candidate memory cells:

[0080]

[0081] Memory unit update:

[0082]

[0083] Output Gate:

[0084] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0085] Hidden state:

[0086] h t =o t tanh(C t )

[0087] Output layer: output prediction results;

[0088] Activation function: Softmax for classification tasks or linear activation for regression tasks;

[0089] S3. Active forest fire intervention control method, through the optimization algorithm to determine the resource allocation plan, the following integer linear programming formula can be used for resource optimization:

[0090]

[0091] Among them, x i,r Indicates that resource i is allocated to region j, C ij is the cost of resource allocation;

[0092] Constraints include the total amount of resources, demand satisfaction, and path restrictions:

[0093]

[0094] Objective function explanation: x i,r is a binary decision variable indicating whether resource r is allocated to region i, x i,r =1 means allocation, x i,r =0 means no allocation;

[0095] c i,r is the unit cost of allocating resource r to region i;

[0096] Constraints,In order to ensure the rationality of the resource allocation plan, the following constraints need to be met:

[0097] The total amount of resources is constrained. The total amount of resources cannot exceed its available amount. Let R be the total amount of resources, R i is the resource demand of region i:

[0098]

[0099] Requirements satisfy constraints. The requirements of each constraint must be satisfied. Let D i is the demand in area i:

[0100]

[0101] Path restriction constraints: some resources may only be allocated through specific paths. For example, set resource r to be allocated only to region i through path p:

[0102]

[0103] Non-negativity constraints, decision variables can only be 0 or 1:

[0104] x i,r ∈{0.1}

[0105] S4. Effectively utilize weather conditions for artificial rainfall: Establish close cooperation with the meteorological department, monitor weather changes in real time, especially key parameters such as cloud thickness and humidity, provide accurate data support for artificial rainfall, and apply for and implement artificial rainfall at the critical moment of fire spread according to weather conditions and fire situation, including determining the area, time, intensity, etc. of rainfall to reduce the spread speed and intensity of the fire; By designing a reasonable reward function, the reward function comprehensively considers multiple factors such as fire control effect, resource consumption, and time cost. The reward function can be defined as:

[0106] R (s,a) =ControlEffect(s,a)-ResourceCost(a)-TimeCost(a)

[0107] Among them, ControlEffect(s,a) represents the effect of intervention action a on fire control, ResourceCost(a) represents the resource consumption cost of executing action a, and TimeCost(a) represents the time cost required to execute action a. By maximizing the cumulative reward, the agent can learn the optimal intervention strategy;

[0108] In the actual intervention process, the agent continuously interacts with the environment, collects new experience data, and updates the weights of the policy network through online learning and optimization algorithms. The update process uses the deep Q network (DQN) or policy gradient method. In DQN, the core of the policy update is the Q value update formula, which is defined as follows:

[0109]

[0110] α: learning rate, which controls the step size of network weight update;

[0111] r: the immediate reward obtained by executing action a in state s;

[0112] γ: discount factor, used to weigh the impact of future rewards;

[0113] s′: the next state after executing action a;

[0114] a′: all possible actions in the next state s′;

[0115] The intelligent agent can autonomously learn and adjust according to real-time feedback and intervention effects, and can autonomously adapt to the latest changes in the fire environment; when the fire changes suddenly or new resources are added, the intelligent agent can respond quickly and re-plan the optimal action to ensure the effectiveness of the intervention strategy; finally, by executing and monitoring the intervention strategy.

[0116] (III) Beneficial effects

[0117] The present invention provides a method and system for intelligent prediction and active intervention control of forest fire spread. It has the following beneficial effects:

[0118] 1. In this invention, a deep reinforcement learning algorithm is introduced in a breakthrough way, and it is deeply integrated with a hybrid deep learning model (CNN+LSTM). This innovation not only gives the model the ability to effectively learn and accurately predict in a high-dimensional state space, but also enables it to autonomously learn and adapt to the complexity and uncertainty of forest fire spread. Through the hybrid deep learning model, the present invention realizes the organic integration of spatial and temporal features, fully reveals the complex mechanism of forest fire spread, and significantly improves the accuracy and reliability of prediction.

[0119] 2. In the present invention, based on the intelligent prediction results, the present invention can accurately delineate high-risk areas for forest fire spread and formulate diversified active intervention control strategies, aiming to take measures in advance to control the spread of fire and reduce losses. At the same time, during the intervention implementation process, the system monitors and evaluates the intervention effect in real time, and dynamically adjusts the intervention strategy and resource allocation plan based on the feedback results. This dynamic adjustment mechanism ensures the effectiveness and flexibility of the intervention measures and provides strong support for forest fire prevention and control. BRIEF DESCRIPTION OF THE DRAWINGS

[0120] Figure 1 It is a schematic diagram used to show the various components of the system and their interrelationships in the present invention. DETAILED DESCRIPTION

[0121] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0122] Embodiment 1:

[0123] like Figure 1As shown, an embodiment of the present invention provides a forest fire spread intelligent prediction and active intervention system, including: a data acquisition module, a data preprocessing module, an intelligent prediction module, a model optimization and verification module, an active intervention decision module, an intervention strategy optimization module, a real-time monitoring and evaluation module, a user interaction interface module and a communication module.

[0124] A control method for a forest fire spread intelligent prediction and active intervention system comprises the following steps:

[0125] S1. Meteorological data acquisition: real-time and historical meteorological data including temperature, humidity, wind speed, wind direction and precipitation are acquired through meteorological stations and satellite remote sensing. Topographic and geomorphological acquisition: topographic maps and DEM digital elevation model data of the region are acquired using GIS system to analyze the characteristics of altitude, slope, aspect and soil type. Vegetation distribution: vegetation type, density, moisture content and flammability data are acquired by combining satellite remote sensing images and ground surveys. Historical forest fire records: detailed records of historical fires are acquired from forestry departments or relevant institutions, including the location of the fire point, the path of spread, the intensity of the fire and the time of extinguishing. Human activities: population density, road distribution and fire source management records are evaluated through population census and transportation network data.

[0126] S2. Data collection and preprocessing,For data with a distribution close to normal, the standard deviation range is used for outlier detection. The calculation process is as follows:

[0127] To calculate the mean and standard deviation, first calculate the mean μ and standard deviation σ of the data set:

[0128]

[0129]

[0130] Where N is the number of values ​​in the data set, x i is the i-th value in the data set;

[0131] Noise removal, missing value filling and standardization processing. First, the input layer receives the preprocessed raw data sequence, which can be a time series signal, such as an audio signal or sensor data; next, the data passes through the first convolution layer, which uses 32 3x3 convolution kernels to extract local spatial features, and then the data is reduced in dimension through the maximum pooling layer to reduce the redundant information of the features; the second convolution layer uses 64 3x3 convolution kernels to further extract higher-level spatial features, and the pooling layer is also used to reduce the feature dimension; the feature maps processed by the convolution layer and the pooling layer are flattened and input to the LSTM layer. The LSTM layer consists of two stacked LSTM layers, each containing 128 units, responsible for capturing time Dynamic features in the sequence; Subsequently, these features are further processed by the fully connected layer to generate the \(k\)th noise pattern \(P_k(t)\) and the corresponding weight coefficient \(\alpha_k\). Finally, the output layer applies these noise patterns and weight coefficients to calculate the denoised data values; During the training process, we use a standardized historical data set as the training set and adopt the mean square error MSE as the loss function. The model parameters are adjusted through the back propagation algorithm to minimize the error between the denoised data and the actual data; This deep learning model that combines CNN and LSTM can effectively and adaptively identify and remove noise in the data, thereby significantly improving the quality and reliability of the data;

[0132] The noise removal formula is:

[0133]

[0134] in, is the data value after noise removal at time point t; X t is the value of the original data at time point t; M i (t) is the value of the i-th noise pattern at time point t, generated by the deep learning model; W i is the corresponding weight coefficient, which is learned through the training process; N is the number of noise patterns;

[0135] Noise Mode M i The calculation of (t) depends on the output of the deep learning model. Assuming that the noise pattern is generated by the features predicted by the model, it can be expressed as:

[0136] M i (t) = f i (CNN(X t )·RNN)

[0137] Among them, CNN (X t ) is the original data X t The features extracted by the convolutional layer are applied; RNN is the time series features processed by the LSTM layer; fi (·) is the i-th noise pattern generated by the fully connected layer or other output layers;

[0138] Missing value filling adopts a new spatiotemporal interpolation method, combines time series analysis and spatial interpolation technology, integrates time and space features to predict missing values, and builds a spatiotemporal model to predict and fill missing data points. The specific filling formula is:

[0139] FilledValue t·j =f(TimeScriesComponent t , SpatialComponent j )

[0140] Among them, FilledValue t·j represents the filling value at time point t and spatial position j, and f is the spatiotemporal interpolation function, which comprehensively considers the temporal continuity and spatial correlation of the data; TimeScriesComponent t Represents the value of the time series component at time point t, obtained through time series analysis: SpatialComponent j Represents the value of the spatial component at spatial position j, obtained by spatial interpolation technology;

[0141] The calculation formula for meteorological data is:

[0142]

[0143] in, is the fill value, temperature, humidity at time point t and spatial position s; T t is the time series component, representing the meteorological data and temperature at time point t; H s is the spatial component, representing the meteorological data and humidity at the spatial location s; W s are other spatial features, such as wind speed and wind direction; α T , α H , α W , β and δ are model parameters;

[0144] The calculation formula for topography is:

[0145]

[0146] in, is the filled value, soil moisture, at time point t and spatial location s; E S is the spatial component, representing the altitude data of the spatial position s; S s is the spatial component, indicating the slope of the spatial position s; A sis the spatial component, indicating the slope direction of the spatial position s; α E , α s , α A , β and δ are model parameters;

[0147] The calculation formula for vegetation distribution is:

[0148]

[0149] in, is the filling value at time point t and spatial position s, the vegetation density; V s is the spatial component, indicating the vegetation type and density at the spatial location s; D s is the other spatial component, water content; α v , α D , β and δ are model parameters;

[0150] The calculation formula for historical forest fire records is:

[0151]

[0152] in, is the fill value at time point t and spatial location s, fire intensity; F s is the spatial component, representing the historical fire record and fire intensity at the spatial location s; P s is the other spatial component, the location of the fire point; α F , α P , β and δ are model parameters;

[0153] The calculation formula for human activities is:

[0154]

[0155] in, is the fill value at time point t and spatial location s, population density; P s is the spatial component, representing the population density at spatial location s; R s is the other spatial component, road distribution; α p , α R , β and δ are model parameters;

[0156] According to the actual distribution and characteristics of the data, the normalization parameters are adjusted dynamically. The specific normalization formula is:

[0157]

[0158] Among them, StandardizedData t represents the standardized data at time point t, μ t and σ′ tRespectively represent the mean of the data at time point t and the dynamically adjusted standard deviation, adaptive adjustment strategy:

[0159] Threshold setting: Set the threshold for data distribution changes. When the changes in the real-time mean and standard deviation exceed the set threshold, dynamic adjustment is triggered;

[0160] Adjustment factor: Use an adaptive adjustment factor to update the normalized parameters, which can be calculated based on the changes in historical data;

[0161] The dynamic adjustment formula is:

[0162] μ′ t =μ t-1 +β·(μ t -μ t-1 )

[0163] σ′ t =σ t-1 +β·(σ t -σ t-1 )

[0164] Among them, μ′ t and σ′ t is the adjusted mean and standard deviation, β is the adjustment factor, usually between 0 and 1, which determines the weight of historical and current data on the adjustment. In the data preprocessing stage, the original data is first standardized to ensure that all data have the same dimension and range, thereby improving the stability and efficiency of model training. Then, the unnecessary noise in the data is removed through the denoising process to ensure that the model can focus on useful signals. Then, the processed data is divided into a training set and a test set, where the training set is used for model training and the test set is used to evaluate the model performance. The following feature extraction process includes multiple steps:

[0165] First, the data is passed through the first convolutional layer to extract spatial features, which captures the basic local information in the data;

[0166] Then, the data enters the second convolutional layer, which further extracts higher-dimensional spatial features to obtain more complex feature representations. The feature map output by the convolutional layer is processed by the maximum pooling layer. The pooling operation reduces the feature dimension and the computational burden by selecting the most significant features, while retaining important spatial information. Next, the feature map after the pooling layer is flattened and sent to the fully connected layer for further processing and high-level feature fusion; the fully connected layer comprehensively processes the flattened feature map to generate feature representations for subsequent analysis; then, these feature representations are input to the LSTM layer, which processes the output of the fully connected layer, captures dynamic changes in the time series, and identifies the temporal characteristics and patterns of the data; in this way, the model can comprehensively consider the characteristics of both space and time for in-depth analysis and prediction;

[0167] In terms of innovation and optimization of the prediction model, the input layer receives carefully pre-processed multi-dimensional data, which are first subjected to spatial feature extraction through multiple convolutional layers. The convolution operation captures multi-scale spatial features in the data through convolution kernels of different sizes. Next, the pooling layer is used to reduce the dimension and extract the most significant features, reduce the amount of calculation and prevent overfitting. After the convolution and pooling operations, the feature map is flattened and sent to one or more fully connected layers, which are used to further fuse the spatial features and perform high-level abstraction. Next, these features are subjected to time series analysis through the LSTM layer to capture the temporal relationship and dynamic changes in the data.

[0168] Input layer: receives preprocessed multidimensional data, such as time series signals or image data;

[0169] Convolutional layer, CNN part

[0170] First convolutional layer: convolution kernel: size is 3×3, number of layers: 3 convolutional layers;

[0171] formula:

[0172]

[0173] Among them, C1(x,y) is the feature map after the convolution operation, W i,j is the convolution kernel weight, X x+i,y+j is the input data value, b is the bias term;

[0174] Second convolutional layer: convolution kernel: size 5×5,

[0175] formula:

[0176]

[0177] Where C2(x,y) is the feature map after the convolution operation;

[0178] Pooling layer:

[0179] Type: max pooling, pooling size 2×2;

[0180] formula:

[0181]

[0182] Among them, P i,j It is the feature map after pooling;

[0183] Fully connected layers, number of layers: one or more fully connected layers;

[0184] Function: to fuse and abstract the features extracted by the convolution kernel pooling layer at a high level;

[0185] formula:

[0186]

[0187] where F k is the output of the fully connected layer, A i is the activation value of the previous layer, W ik is the weight of the fully connected layer, b k is the bias term;

[0188] LSTM layer, RNN part, number of layers: 2 LSTM layers; number of units: 50 LSTM units per layer;

[0189] formula:

[0190] Forget Gate:

[0191] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0192] Input Gate:

[0193] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0194] Candidate memory cells:

[0195]

[0196] Memory unit update:

[0197]

[0198] Output Gate:

[0199] o t =α(W o ·[h t-1 ,x t ]+b o )

[0200] Hidden state:

[0201] h t =o t tanh(C t )

[0202] Output layer: output prediction results;

[0203] Activation function: Softmax for classification tasks or linear activation for regression tasks;

[0204] S3. Active forest fire intervention control method, through the optimization algorithm to determine the resource allocation plan, the following integer linear programming formula can be used for resource optimization:

[0205]

[0206] Among them, x i,r Indicates that resource i is allocated to region j, C ij is the cost of resource allocation;

[0207] Constraints include the total amount of resources, demand satisfaction, and path restrictions:

[0208]

[0209] Objective function explanation: x i,r is a binary decision variable indicating whether resource r is allocated to region i, x i,r =1 means allocation, x i,r =0 means no allocation;

[0210] c i,r is the unit cost of allocating resource r to region i;

[0211] Constraints,In order to ensure the rationality of the resource allocation plan, the following constraints need to be met:

[0212] The total amount of resources is constrained. The total amount of resources cannot exceed its available amount. Let R be the total amount of resources, R i is the resource demand of region i:

[0213]

[0214] Requirements satisfy constraints. The requirements of each constraint must be satisfied. Let D i is the demand in area i:

[0215]

[0216] Path restriction constraints: some resources may only be allocated through specific paths. For example, set resource r to be allocated only to region i through path p:

[0217]

[0218] Non-negativity constraints, decision variables can only be 0 or 1:

[0219] x i,r ∈{0.1}

[0220] S4. Effectively utilize weather conditions for artificial rainfall: Establish close cooperation with the meteorological department, monitor weather changes in real time, especially key parameters such as cloud thickness and humidity, provide accurate data support for artificial rainfall, and apply for and implement artificial rainfall at the critical moment of fire spread according to weather conditions and fire situation, including determining the area, time, intensity, etc. of rainfall to reduce the spread speed and intensity of the fire; By designing a reasonable reward function, the reward function comprehensively considers multiple factors such as fire control effect, resource consumption, and time cost. The reward function can be defined as:

[0221] R (s,a) =ControlEffedt(s,a)-ResourceCost(a)-TimeCost(a)

[0222] Among them, ControlEffect(s,a) represents the effect of intervention action a on fire control, ResourceCost(a) represents the resource consumption cost of executing action a, and TimeCost(a) represents the time cost required to execute action a. By maximizing the cumulative reward, the agent can learn the optimal intervention strategy;

[0223] In the actual intervention process, the agent continuously interacts with the environment, collects new experience data, and updates the weights of the policy network through online learning and optimization algorithms. The update process uses the deep Q network (DQN) or policy gradient method. In DQN, the core of the policy update is the Q value update formula, which is defined as follows:

[0224]

[0225] α: learning rate, which controls the step size of network weight update;

[0226] r: the immediate reward obtained by executing action a in state s;

[0227] γ: discount factor used to weigh the impact of future rewards;

[0228] s′: the next state after executing action a;

[0229] a′: all possible actions in the next state s′;

[0230] The intelligent agent can autonomously learn and adjust according to real-time feedback and intervention effects, and can autonomously adapt to the latest changes in the fire environment; when the fire changes suddenly or new resources are added, the intelligent agent can respond quickly and re-plan the optimal action to ensure the effectiveness of the intervention strategy; finally, by executing and monitoring the intervention strategy.

[0231] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A forest fire spread intelligent prediction and active intervention system, characterized by: include: Data acquisition module, data preprocessing module, intelligent prediction module, model optimization and verification module, active intervention decision module, intervention strategy optimization module, real-time monitoring and evaluation module, user interaction interface module and communication module.

2. A control method for a forest fire spread intelligent prediction and active intervention system, characterized in that: The following steps are involved: S1. Meteorological data acquisition: real-time and historical meteorological data including temperature, humidity, wind speed, wind direction and precipitation are acquired through meteorological stations and satellite remote sensing. Topographic and geomorphological acquisition: topographic maps and DEM digital elevation model data of the region are acquired using GIS system to analyze the characteristics of altitude, slope, aspect and soil type. Vegetation distribution: vegetation type, density, moisture content and flammability data are acquired by combining satellite remote sensing images and ground surveys. Historical forest fire records: detailed records of historical fires are acquired from forestry departments or relevant institutions, including the location of the fire point, the path of spread, the intensity of the fire and the time of extinguishing. Human activities: population density, road distribution and fire source management records are evaluated through population census and transportation network data. S2. Data collection and preprocessing,For data with a distribution close to normal, the standard deviation range is used for outlier detection. The calculation process is as follows: To calculate the mean and standard deviation, first calculate the mean μ and standard deviation σ of the data set: Where N is the number of values ​​in the data set, x i is the i-th value in the data set; Noise removal, missing value filling and standardization processing. First, the input layer receives the preprocessed raw data sequence, which can be a time series signal, such as an audio signal or sensor data; next, the data passes through the first convolution layer, which uses 32 3x3 convolution kernels to extract local spatial features, and then the data is reduced in dimension through the maximum pooling layer to reduce the redundant information of the features; the second convolution layer uses 64 3x3 convolution kernels to further extract higher-level spatial features, and the pooling layer is also used to reduce the feature dimension; the feature maps processed by the convolution layer and the pooling layer are flattened and input to the LSTM layer. The LSTM layer consists of two stacked LSTM layers, each containing 128 units, responsible for capturing time series The dynamic features in the column are then further processed by the fully connected layer to generate the kth noise pattern Pk(t) and the corresponding weight coefficient alphak. Finally, the output layer applies these noise patterns and weight coefficients to calculate the denoised data values. During the training process, we use the standardized historical data set as the training set and adopt the mean square error (MSE) as the loss function. The model parameters are adjusted through the back propagation algorithm to minimize the error between the denoised data and the actual data. This deep learning model combining CNN and LSTM can effectively and adaptively identify and remove noise in the data, thereby significantly improving the quality and reliability of the data. The noise removal formula is: in, is the data value after removing noise at time point t; X t is the value of the original data at time point t; M i (t) is the value of the i-th noise pattern at time point t, generated by the deep learning model; W i is the corresponding weight coefficient, which is learned through the training process; N is the number of noise patterns; Noise Mode M i The calculation of (t) depends on the output of the deep learning model. Assuming that the noise pattern is generated by the features predicted by the model, it can be expressed as: M i (t)=f i (CNN(X t )·RNN) Among them, CNN (X t ) is the original data X t The features extracted by the convolutional layer are applied; RNN is the time series features processed by the LSTM layer; f i (·) is the i-th noise pattern generated by the fully connected layer or other output layers; Missing value filling adopts a new spatiotemporal interpolation method, combines time series analysis and spatial interpolation technology, integrates time and space features to predict missing values, and builds a spatiotemporal model to predict and fill missing data points. The specific filling formula is: FilledValue t·j =f(TimeScriesComponent t ,SpatialComponent j ) Among them, FilledValue t·j represents the filling value at time point t and spatial position j, and f is the spatiotemporal interpolation function, which comprehensively considers the temporal continuity and spatial correlation of the data; TimeScriesComponent t Represents the value of the time series component at time point t, obtained through time series analysis: SpatialComponent j Represents the value of the spatial component at spatial position j, obtained by spatial interpolation technology; The calculation formula for meteorological data is: in, is the fill value (temperature, humidity) at time point t and spatial location s; T t is the time series component, representing the meteorological data (temperature) at time point t; H s is the spatial component, representing the meteorological data (humidity) at the spatial location s; W s are other spatial features (wind speed, wind direction); α T , α H , α W , β and δ are model parameters; The calculation formula for topography is: in, is the filled value (soil moisture) at time point t and spatial location s; E S is the spatial component, representing the altitude data of the spatial position s; S s is the spatial component, indicating the slope of the spatial position s; A s is the spatial component, indicating the slope direction of the spatial position s; α E , α s , α A , β and δ are model parameters; The calculation formula for vegetation distribution is: in, is the filling value (vegetation density) at time point t and spatial position s; V s is the spatial component, indicating the vegetation type and density at the spatial location s; D s is the other spatial component (water content); α v , α D , β and δ are model parameters; The calculation formula for historical forest fire records is: in, is the fill value (fire intensity) at time point t and spatial location s; F s is the spatial component, representing the historical fire record (fire intensity) at the spatial location s; P s is the other spatial component (fire location); α F , α P , β and δ are model parameters; The calculation formula for human activities is: in, is the fill value (population density) at time point t and spatial location s; P s is the spatial component, representing the population density at spatial location s; R s is other spatial components (road distribution); α p , α R , β and δ are model parameters; According to the actual distribution and characteristics of the data, the normalization parameters are adjusted dynamically. The specific normalization formula is: Among them, StandardizedData t represents the standardized data at time point t, μ t and σ′ t Respectively represent the mean of the data at time point t and the dynamically adjusted standard deviation, adaptive adjustment strategy: Threshold setting: Set the threshold for data distribution changes. When the changes in the real-time mean and standard deviation exceed the set threshold, dynamic adjustment is triggered; Adjustment factor: Use an adaptive adjustment factor to update the normalized parameters, which can be calculated based on the changes in historical data; The dynamic adjustment formula is: m′ t =μ t-1 +β·(μ t -m t-1 ) in t =s t-1 +β·(σ t -s t-1 ) Among them, μ′ t and σ′ t is the adjusted mean and standard deviation, β is the adjustment factor, usually between 0 and 1, which determines the weight of historical and current data on the adjustment. In the data preprocessing stage, the original data is first standardized to ensure that all data have the same dimension and range, thereby improving the stability and efficiency of model training. Then, the unnecessary noise in the data is removed through the denoising process to ensure that the model can focus on useful signals. Then, the processed data is divided into a training set and a test set, where the training set is used for model training and the test set is used to evaluate the model performance. The following feature extraction process includes multiple steps: First, the data is passed through the first convolutional layer to extract spatial features, which captures the basic local information in the data; Then, the data enters the second convolutional layer, which further extracts higher-dimensional spatial features to obtain more complex feature representations. The feature map output by the convolutional layer is processed by the maximum pooling layer. The pooling operation reduces the feature dimension and the computational burden by selecting the most significant features, while retaining important spatial information. Next, the feature map after the pooling layer is flattened and sent to the fully connected layer for further processing and high-level feature fusion; the fully connected layer comprehensively processes the flattened feature map to generate feature representations for subsequent analysis; then, these feature representations are input to the LSTM layer, which processes the output of the fully connected layer, captures dynamic changes in the time series, and identifies the temporal characteristics and patterns of the data; in this way, the model can comprehensively consider the characteristics of both space and time for in-depth analysis and prediction; In terms of innovation and optimization of the prediction model, the input layer receives carefully pre-processed multi-dimensional data, which are first subjected to spatial feature extraction through multiple convolutional layers. The convolution operation captures multi-scale spatial features in the data through convolution kernels of different sizes. Next, the pooling layer is used to reduce the dimension and extract the most significant features, reduce the amount of calculation and prevent overfitting. After the convolution and pooling operations, the feature map is flattened and sent to one or more fully connected layers, which are used to further fuse the spatial features and perform high-level abstraction. Next, these features are subjected to time series analysis through the LSTM layer to capture the temporal relationship and dynamic changes in the data. Input layer: receives preprocessed multidimensional data (such as time series signals or image data); Convolutional layer (CNN part) First convolutional layer: convolution kernel: size is 3×3, number of layers: 3 convolutional layers; formula: Among them, C1(x, y) is the feature map after the convolution operation, W i，j is the convolution kernel weight, X x+i,y+j is the input data value, b is the bias term; The second convolution layer: convolution kernel: size 5×5, formula: Where C2(x, y) is the feature map after the convolution operation; Pooling layer: Type: max pooling, pooling size 2×2; formula: Among them, P i，j It is the feature map after pooling; Fully connected layers, number of layers: one or more fully connected layers; Function: to fuse and abstract the features extracted by the convolution kernel pooling layer at a high level; formula: where F k is the output of the fully connected layer, A i is the activation value of the previous layer, W ik is the weight of the fully connected layer, b k is the bias term; LSTM layer (RNN part), number of layers: 2 LSTM layers; number of units: 50 LSTM units per layer; formula: Forget Gate: f t =σ(W f ·[h t-1 ,x t ]+b f ) Input Gate: i t =σ(W i ·[h t-1 ,x t ]+b i ) Candidate memory cells: Memory unit update: Output Gate: the t =σ(W o ·[h t-1 ,x t ]+b o ) Hidden state: h t =o t ·tanh(C t ) Output layer: output prediction results; Activation function: Softmax (for classification tasks) or linear activation (for regression tasks); S3. Active forest fire intervention control method, through the optimization algorithm to determine the resource allocation plan, the following integer linear programming formula can be used for resource optimization: Among them, x i，r Indicates that resource i is allocated to region j, C ij is the cost of resource allocation; Constraints include the total amount of resources, demand satisfaction, and path restrictions: Objective function explanation: x i，r is a binary decision variable indicating whether resource r is allocated to region i (x i，r =1 means allocation, x i，r =0 means no allocation); c i，r is the unit cost of allocating resource r to region i; Constraints,In order to ensure the rationality of the resource allocation plan, the following constraints need to be met: The total amount of resources is constrained. The total amount of resources cannot exceed its available amount. Let R be the total amount of resources, R i is the resource demand of region i: Requirements satisfy constraints. The requirements of each constraint must be satisfied. Let D i is the demand in area i: Path restriction constraints: some resources may only be allocated through specific paths. For example, set resource r to be allocated only to region i through path p: Non-negativity constraints, decision variables can only be 0 or 1: x i，r ∈{0.1} S4. Effectively utilize weather conditions for artificial rainfall: Establish close cooperation with the meteorological department, monitor weather changes in real time, especially key parameters such as cloud thickness and humidity, provide accurate data support for artificial rainfall, and apply for and implement artificial rainfall at the critical moment of fire spread according to weather conditions and fire situation, including determining the area, time, intensity, etc. of rainfall to reduce the spread speed and intensity of the fire; By designing a reasonable reward function, the reward function comprehensively considers multiple factors such as fire control effect, resource consumption, and time cost. The reward function can be defined as: R(s,a)=ControlEffect(s,a)-ResourceCost(a)-TimeCost(a) Among them, ControlEffect(s, a) represents the effect of intervention action a on fire control, ResourceCost(a) represents the resource consumption cost of executing action a, and TimeCost(a) represents the time cost required to execute action a. By maximizing the cumulative reward, the agent can learn the optimal intervention strategy; In the actual intervention process, the agent continuously interacts with the environment, collects new experience data, and updates the weights of the policy network through online learning and optimization algorithms. The update process uses the deep Q network (DQN) or policy gradient method. In DQN, the core of the policy update is the Q value update formula, which is defined as follows: α: learning rate, which controls the step size of network weight update; r: the immediate reward obtained by executing action a in state s; γ: discount factor used to weigh the impact of future rewards; s′: the next state after executing action a; a′: all possible actions in the next state s′; The intelligent agent can autonomously learn and adjust according to real-time feedback and intervention effects, and can autonomously adapt to the latest changes in the fire environment; when the fire changes suddenly or new resources are added, the intelligent agent can respond quickly and re-plan the optimal action to ensure the effectiveness of the intervention strategy; finally, by executing and monitoring the intervention strategy.

Images