A forest fire prediction method based on digital twins

Forest fire prediction through digital twin technology, and time series decomposition and deep learning models are used to solve the problems of low accuracy and difficulty in detection in the existing technology, and more accurate fire source positioning and fire spread paths are achieved. Prediction provides effective strategies for rescue.

CN118627936BActive Publication Date: 2024-11-15CHENGDU SENTUO MURAN TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411087637.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-09
Publication Date
2024-11-15
Estimated Expiration
2044-08-09

AI Technical Summary

Technical Problem

The existing forest fire prediction methods rely on experience, have low accuracy, and are difficult to detect small fire sources that are blocked, and fail to effectively consider the impact of human activities, resulting in low prediction accuracy.

Method used

Using a forest fire prevention prediction method based on digital twins, the seasonal decomposition and SSA singular spectrum analysis are performed by obtaining historical fire time series data, combining stacked LSTM neural network and full connection layer to generate a digital twin database and a three-dimensional visualization platform to realize the fire source Positioning and fire spread path prediction.

Benefits of technology

Improve the accuracy and accuracy of forest fire prediction, can monitor forest state in real time in virtual space, provide correct strategy guidance, reduce dependence on real-time data, avoid gradient vanishing and explosion problems, and enhance the generalization ability of the model .

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118627936B_ABST
    Figure CN118627936B_ABST
Patent Text Reader

Abstract

The present application provides a forest fire prediction method based on digital twins, which belongs to the field of forest fire prevention. The method obtains time series data corresponding to historical fires in a forest to be tested, performs seasonal decomposition and SSA singular spectrum analysis on the time series data, imports the time series data into a stacked LSTM neural network layer, adds a fully connected layer after the LSTM neural network layer, obtains an optimal parameter model through parameter adjustment and training, then obtains fire data and vegetation data of the forest to be tested to generate a digital twin database, transmits data between the digital twin database and a three-dimensional forest visualization platform to generate a digital twin model of the forest to be tested, generates fire alarm information when a fire is detected based on the fire data, obtains the current time series data, and uses the optimal parameter model to locate the fire source and predict the fire development trend, thereby realizing timely and accurate monitoring of forest fires and accurate prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of forest fire prevention, and in particular to a forest fire prediction method based on digital twins. Background Art

[0002] Forest fire is a natural disaster that has great destructive power on the environment, ecology, economy and many other aspects. People all over the world hope to improve the ability to predict forest fires. Forest fire prediction is essentially a process of nonlinear fitting of the probability of fire occurrence based on a series of parameters, such as temperature, humidity, wind direction, wind force, vegetation distribution, etc., to conduct risk assessment (such as the Chinese patent "A forest fire early warning model and system based on deep learning technology", application number: 202010478816.0), but the above prediction method relies too much on experience, so the prediction accuracy is not high. In addition, there are also articles in the prior art that use infrared thermal imaging combined with visible light imaging to warn forest fires in the initial stage of occurrence (such as the Chinese patent "Forest fire detection method based on neural network", application number: 202010701658.0), but the above method does not consider the impact of human activities on imaging, and due to the complex terrain of the forest, it is difficult to detect small fire sources that are blocked. Therefore, it is still necessary to improve the prediction accuracy of forest fire prevention.

[0003] Digital twin technology refers to the full use of physical models, sensors, operation history and other data, integrating multi-disciplinary, multi-physical quantity, multi-scale, and multi-probability simulation processes, and completing mapping in virtual space, thereby reflecting the entire life cycle of the corresponding physical equipment. It can be regarded as a digital mapping system of one or more important and interdependent equipment systems. Digital twin technology has been increasingly used in various application fields such as risk identification, disaster prediction, and failure prediction. However, the application of this technology in the field of forest fire prevention is still relatively small. How to establish a digital twin system in the field of forest fire prevention and improve its prediction accuracy is of great significance for improving the prediction effect of forest fire prevention and providing guidance strategies for subsequent disaster relief. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the present invention proposes a forest fire prediction method based on digital twins.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A forest fire prevention prediction method based on digital twins, the method comprising the following steps:

[0007] S1: Obtain the time series data corresponding to the historical fires of the forest to be tested, perform seasonal decomposition on the time series data, decompose the time series data into seasonal cycle components, trend components and random components, and perform SSA singular spectrum analysis on the decomposed random components;

[0008] S2: The seasonal component, trend component and random component of the time series data processed by SSA singular spectrum analysis are imported into the stacked LSTM neural network layer, and a fully connected layer is added after the LSTM neural network layer to obtain the optimal parameter model through parameter adjustment and training;

[0009] S3: Obtain fire protection data, vegetation data, meteorological data, and terrain data of the forest to be tested to generate a digital twin database, and transmit data between the digital twin database and the three-dimensional forest visualization platform to generate a digital twin model of the forest to be tested;

[0010] S4: When a fire is detected based on fire data, fire alarm information is generated, current time series data is acquired, the optimal parameter model is used to locate the fire source and predict the fire development trend.

[0011] Furthermore, in step S1, time series data is obtained by analyzing video images of the fire.

[0012] Furthermore, in step S1, the X-11 time series decomposition method is used to perform seasonal decomposition on the time series data, using the additive model formula:

[0013] ;

[0014] In the formula, represents time series data, represents the seasonal cycle component, represents the trend component, Represents the random component.

[0015] Furthermore, in step S2, a stacked LSTM neural network layer is used to extract feature information of the seasonal cycle component, the trend component and the random component processed by SSA singular spectrum analysis, and the data analysis is completed through a fully connected layer to generate the final output.

[0016] Furthermore, in step S2, Bayesian parameter adjustment is used to obtain the optimal parameters.

[0017] Furthermore, in step S3, the accuracy of the model is evaluated by the mean square error and the mean absolute error percentage.

[0018] Furthermore, the fire protection data includes data collected by temperature sensors, smoke sensors, flame detectors, and gas sensors.

[0019] Furthermore, vegetation data include forest coverage and forest stand structure, meteorological data include temperature, air humidity, soil humidity, wind direction and wind speed, and terrain data include slope, slope direction, slope position and power transmission and distribution lines.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] 1. By building a digital twin model of the forest, based on current big data processing, Internet of Things, cloud computing, artificial intelligence, and virtual reality technology, the state of the forest can be displayed in three dimensions and updated to the latest state. For back-end monitoring personnel, forest information such as current temperature and humidity, and fire-susceptible points can be obtained without on-site monitoring and investigation. It can also be used in fields such as forest weather forecasts based on real-time status updates, with a user-friendly interface and strong interactivity;

[0022] 2. The present invention adopts a combination of X-11 seasonal decomposition method and stacked LSTM neural network model. After decomposing the time series data using the X-11 seasonal decomposition method, the irregular random components are decomposed using SSA singular spectrum analysis to smooth the random components and reduce the prediction difficulty of the model. The stacked LSTM is used to better capture the long-term dependencies in the time series data, avoiding the need to rely on real-time data, and avoiding the gradient vanishing and gradient explosion problems in the RNN, so as to more accurately predict the relevant situations of future fires based on historical fire information, such as locating the fire source and predicting the fire spread path, and can further provide correct strategic guidance for rescue, such as guiding the fire escape route;

[0023] 3. The present invention also adds a fully connected layer to the neural network model, and uses the fully connected layer to analyze data, which is conducive to the model learning the nonlinear relationship between features, so that the model can better fit the training data and have better generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Flowchart of a forest fire prediction method based on digital twins in an embodiment of the present invention;

[0025] Figure 2 Schematic diagram of the working principle of the neural network model in an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0027] See also Figure 1 to Figure 2, the present application provides a forest fire prediction method based on digital twins. In this method, when obtaining the fire data of the monitoring points of the forest to be tested, the forest to be tested can select forest areas that are prone to fire and need to be focused on. These areas have more historical fire-related data, which is conducive to the selection of more parameters and data when establishing a digital twin database in the future. Taking Sichuan Province as an example, the forest coverage rate of the province is about 35%-38%. Due to its special geographical and climatic conditions, Sichuan Province is a province with frequent forest and grassland fires, and it is also a key area for forest and grassland fire prevention in the country. According to statistical data, there are 175 counties (cities and districts) involved in forest and grassland fire prevention tasks in the province, of which 35 and 81 are high-risk areas and high-risk areas for forest fires recognized by the state. Especially in Panzhihua, Liangshan, Ganzi and other places, due to the high frequency, wide range and high intensity of extreme weather, the forest and grassland fire risk meteorological level can reach level 5, and the situation is very serious.

[0028] In order to realize forest fire prevention prediction, a suitable prediction model needs to be selected. In the present invention, a stacked LSTM neural network model is selected to realize fire prediction. In order to improve the accuracy of prediction, an X-11 seasonal decomposition method is added to process time series data, and a fully connected layer is added before the output layer of the neural network model to improve the generalization ability and information processing ability of the model.

[0029] First, in the first step, the time series data corresponding to the historical fires of the forest to be tested is obtained by analyzing the video images of the historical fires, and the time series data is seasonally decomposed, the time series data is decomposed into seasonal cycle components, trend components and random components, and the decomposed random components are processed by SSA singular spectrum analysis.

[0030] The seasonal decomposition method in the present invention adopts the X-11 time series decomposition method. The X-11 time series decomposition method is a classic time series analysis method, which can decompose the original time series data Xt into the following parts:

[0031] Seasonal cycle component (C t ): Also known as periodic term or seasonal term, it reflects various external factors with a specific period as the cycle;

[0032] Trend component (T t ): Also known as the trend item, it reflects the overall trend of time series data over a long period of time;

[0033] Random component (R t ): Also known as the residual term, it reflects the impact of changes in accidental factors other than the seasonal cycle component and the trend component on the original data.

[0034] There are two types of decomposition models: additive model and multiplicative model. The present invention adopts additive model to realize data decomposition. The additive model formula is:

[0035] ;

[0036] In the formula, represents time series data, represents the seasonal cycle component, represents the trend component, Represents the random component.

[0037] First, use the central moving average method to extract the original time series data X t Separate the trend component T t , by generating a new smoothed series value corresponding to the center point of the averaged time series data. t After separation, convert using the additive model:

[0038] ;

[0039] That is, the original time series data X t Subtract the trend component T t Then we get the seasonal cycle component C t and the random component R t . Continuing to perform moving average on the remaining part, the random component Rt can be removed to obtain the seasonal periodic component Ct.

[0040] The above-mentioned X-11 time series decomposition method can be used to decompose complex time series data. The decomposed seasonal cycle components and trend components make it easier for the stacked LSTM model to analyze and extract features from time series data, greatly reducing the difficulty of time series prediction.

[0041] Since the random component R t Irregular, time information features are difficult to capture, so the random component R t First perform SSA processing and then substitute it into the stacked LSTM model. SSA processing is a decomposition and reconstruction method that separates the signal with the smallest contribution rate to smooth the sequence of random components, thereby reducing the prediction difficulty of the stacked LSTM model. The above decomposition is achieved by calling the seas() function in R language.

[0042] In the second step, the seasonal cycle component, trend component and random component processed by SSA singular spectrum analysis of the time series data are imported into the stacked LSTM neural network layer, and a fully connected layer is added after the LSTM neural network layer to obtain the optimal parameter model through parameter adjustment and training. Specifically, the stacked LSTM neural network layer is used to extract feature information of the seasonal cycle component, trend component and random component processed by SSA singular spectrum analysis, and the data is analyzed through the fully connected layer to generate the final output. Bayesian parameter adjustment is used to obtain the optimal parameters, and the accuracy of the model is evaluated by the mean square error and mean absolute error percentage.

[0043] Stacked LSTM is a deep learning neural network model, a special recurrent neural network (RNN) with memory units and gating mechanisms. It avoids the gradient vanishing and gradient exploding problems in traditional RNNs, can effectively capture long-term dependencies, and is used to model and predict sequence data. Stacked LSTM is generated through multiple layers of LSTM, and the sequence output of the upper layer is input to the lower layer. Compared with the traditional LSTM neural network, this structure increases the depth of the neural network by adding the number of hidden layers, so it has better feature extraction capabilities. The optimal number of LSTM units is determined through training, and finally the regression analysis of the data is completed through the fully connected layer to improve the model's ability to learn nonlinear relationships between features, so that the model can better fit the training data and have better generalization capabilities.

[0044] After the model is established, the data of the training set is input and the parameters are adjusted during the training process. In this invention, Bayesian parameter adjustment is adopted, and the optimal model is saved after the training is completed. Model accuracy refers to the comparison between the training results and the actual values ​​of the model during the training process, and the mean square error, mean absolute error percentage, etc. are used to measure the prediction results.

[0045] The formula for Root Mean Square Error (RMSE) is:

[0046]

[0047] Where n represents the number of samples, is the true value, is the predicted value.

[0048] The formula for Mean Absolute Percent Error (MAPE) is:

[0049]

[0050] Where n represents the number of samples, where n represents the number of samples, is the true value, is the predicted value.

[0051] The data of the training set is obtained through the video images corresponding to the historical fires. The video images are divided into multiple frames, and grayscale processing is performed to obtain serialized input data. The model training is completed by comparing the output value with the actual value until the model accuracy meets the requirements. In the training process of the model, we used a variety of neural network models, such as LSTM and RNN, to compare with the stacked LSTM. The comparison found that the stacked LSTM has a better prediction effect due to the deeper feature level extracted, and solves the gradient disappearance and gradient explosion problems of RNN. In addition, the increase of the fully connected layer can enhance the fitting ability of the model. In the present invention, two LSTM layers are stacked, and a fully connected layer is used to fit the data after stacking for output.

[0052] The third step is to obtain the fire protection data, vegetation data, meteorological data, and terrain data of the forest to be tested to generate a digital twin database, and transmit the data of the digital twin database to the three-dimensional forest visualization platform to generate a digital twin model of the forest to be tested.

[0053] Through the data of GIS system and satellite remote sensing data, combined with virtual reality technology, big data computing and computer graphics, a three-dimensional forest visualization platform is constructed in a 1:1 ratio. The construction of the three-dimensional forest visualization platform belongs to the prior art. This application aims to use the information interaction between the platform and the digital twin database to build a digital twin model of the forest. Among them, the data of the digital twin database includes fire data, vegetation data, meteorological data, terrain data, etc. The collection of fire data is to display fire-susceptible points in the digital twin model and monitor the occurrence of fire. Fire data includes data collected by temperature sensors, smoke sensors, flame detectors, and gas sensors. The layout of these sensors can be selected according to the experience of technicians. It is arranged at points prone to fire, such as around power transmission and distribution lines and areas prone to fire in the past. A mesh layout is required to comprehensively collect and monitor fire conditions. Vegetation data includes forest coverage and forest stand structure, meteorological data includes temperature, air humidity, soil humidity, wind direction, wind speed, and terrain data includes slope, slope direction, slope position, and power transmission and distribution lines. The above data have different accuracy requirements and update timeliness requirements when establishing a forest digital twin model. Among them, meteorological data has high timeliness and high precision requirements to dynamically reflect the actual situation of the forest. Therefore, more sensors can be deployed and the sensor collection frequency is also higher. Vegetation data and terrain data are relatively stable and can be obtained through drones, remote sensing, radar, etc. Considering the difficulty of actual modeling, vegetation data and terrain data can be simplified to a certain extent, and the focus is on extracting data from key areas. For example, power transmission and distribution lines are areas prone to fire, and more data can be collected in areas around the transmission and distribution lines. The terrain also has an important impact on the occurrence and spread of fire, and more data can be collected in areas with large terrain changes.

[0054] Vegetation data, meteorological data, terrain data and fire data are jointly constructed into a digital twin database, which is transmitted to the three-dimensional forest visualization platform in real time, providing data support for the digital twin model of the forest to achieve integration and dynamic display, which is conducive to the technicians to know the situation of the forest in a timely and accurate manner, and can be used for meteorological monitoring and forecasting, real-time monitoring of fire-susceptible points, etc. Through the establishment of the forest digital twin model, the present invention can synchronously update the forest dynamics, realize intelligent monitoring and management of forestry through a combination of virtual and real, and has a friendly interface and strong interactivity. In addition, in addition to being used for forest fire prevention, it can also provide information and guidance for many other purposes such as weather forecasting and afforestation.

[0055] The fourth step is to generate fire alarm information when a fire is detected based on fire data, obtain the current time series data, and use the optimal parameter model to locate the fire source and predict the fire development trend.

[0056] Using the trained optimal parameter model mentioned above to predict fires, we can use the powerful nonlinear mapping ability of neural networks to predict the development of fires in the next period of time based on historical data. The trained neural network is like the human brain, with prediction and analysis functions. It can learn based on historical data. When a fire is detected, it will first issue a fire alarm, and at the same time predict the next fire development trend based on the collection of real-time data, such as predicting the fire spread path, smoke spread path, locating the fire source, and even providing strategic guidance for the next fire rescue and personnel evacuation. The alarm and prediction results are intuitively displayed to the background technicians in the three-dimensional digital twin model, realizing smart forestry prediction and analysis through visualization and human-computer interaction.

[0057] The sensors and detectors used to collect fire data determine the occurrence of fire by monitoring the changes in environmental data when a fire occurs, such as temperature increase, carbon monoxide gas concentration increase, oxygen concentration decrease, smoke and flame appearance. In the present invention, DS18B20 temperature sensor, MQ-2 smoke sensor, SOC embedded flame detector, O2-A2 oxygen sensor, CO-D4 carbon monoxide sensor are selected for detection, and the occurrence of fire is jointly determined by the collection of multiple parameters, among which the appearance of flame and smoke is the main sign of fire occurrence. After the above signals are collected, they can also be combined with temperature and gas data to achieve accurate and timely fire identification to prevent the occurrence of false alarms.

[0058] When a fire is detected, real-time time series data is obtained through the collection of real-time data, such as remote sensing imaging, infrared thermal imaging, ground observation, and video images taken by drones. It should be noted that when constructing time series data, video images taken by the same device should be used to ensure that the size and background of the video images are the same. At the same time, the images during training and the images during prediction are also of the same type, that is, the two fires have similar meteorological environments and vegetation conditions, in order to ensure the accuracy of the prediction results.

[0059] The prediction process of LSTM is as follows: first decompose the video image frame by frame, grayscale the decomposed image to obtain a gray value matrix, sort the gray value matrix by time to obtain a multivariate time series, and use LSTM to perform prediction after seasonal decomposition and SSA processing to obtain the prediction result in the form of multivariate time series-gray value matrix-grayscale image. The prediction result is compared with the grayscale image of the next frame to train and correct the parameters of the neural network until convergence, that is, the neural network reaches the training accuracy. At this time, it is determined that the neural network model can perform the prediction function when an actual fire occurs.

[0060] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0061] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0062] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0063] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A forest fire prediction method based on digital twins, characterized in that: The method comprises the following steps: S1: The time series data corresponding to the historical fires of the forest to be tested is obtained by analyzing the video images of the historical fires, and the X-11 time series decomposition method is used to seasonally decompose the time series data. The X-11 time series decomposition method uses an additive model formula to decompose the time series data into a seasonal cycle component, a trend component and a random component. The additive model formula is: ; In the formula, represents time series data, represents the seasonal cycle component, represents the trend component, represents the random component, After seasonal decomposition, the decomposed random components are processed by SSA singular spectrum analysis; The time series data is a multivariate time series, which is obtained by the following steps: firstly, decomposing the video image frame by frame, graying the decomposed image to obtain a gray value matrix, and sorting the gray value matrix according to time to obtain a multivariate time series; S2: The seasonal cycle component, trend component and random component processed by SSA singular spectrum analysis of the time series data are introduced into the stacked LSTM neural network layer, and a fully connected layer is added after the LSTM neural network layer to obtain the optimal parameter model through parameter adjustment and training; the fully connected layer is located after the stacked LSTM neural network layer and is used to complete the regression analysis of the data to improve the ability of the model to learn the nonlinear relationship between features; S3: Obtain firefighting data, vegetation data, meteorological data, and terrain data of the forest to be tested to generate a digital twin database, and transmit data between the digital twin database and the three-dimensional forest visualization platform to generate a digital twin model of the forest to be tested, wherein the firefighting data includes data collected by temperature sensors, smoke sensors, flame detectors, and gas sensors; the vegetation data includes forest coverage and forest stand structure; the meteorological data includes temperature, air humidity, soil humidity, wind direction, and wind speed; and the terrain data includes slope, slope direction, slope position, and power transmission and distribution lines; S4: When a fire is detected based on fire data, fire alarm information is generated, current time series data is acquired, the optimal parameter model is used to locate the fire source and predict the fire development trend.

2. The method according to claim 1, characterized in that In step S2, a stacked LSTM neural network layer is used to extract feature information from the seasonal cycle component, trend component, and random component processed by SSA singular spectrum analysis, and the data analysis is completed through a fully connected layer to generate the final output.

3. The method according to claim 2, characterized in that In step S2, Bayesian parameter adjustment is used to obtain the optimal parameters.

4. The method according to claim 3, characterized in that In step S3, the accuracy of the model is evaluated by the mean square error and the mean absolute error percentage.