Fire spreading prediction method based on time-space point process

Abstract

The invention discloses a fire spreading prediction method based on a time-space point process, and the method comprises the steps: extracting multi-dimensional features through observation of each time point, and enabling the features to be embedded into a ConvLSTM network, and forming time-space sequence data of fire spreading; local spatio-temporal features are extracted through a ConvLSTM network, and local dependency capture is achieved; a self-attention mechanism is introduced, the dependency relationship between grids is calculated, fire behavior interaction in a remote area is carried out, and remote dependency capture is achieved; and after processing of the ConvLSTM network and the self-attention mechanism, outputting a prediction result of the fire behavior at the future moment. According to the method, the local fire dynamic state is captured through the convolutional long-short-term memory network, and the remote space-time correlation is processed through the self-attention mechanism, so that the prediction precision of fire spreading is improved; the problems that in the prior art, local and global space-time dependence capture is insufficient, the model generalization ability is poor, the calculation cost is high, and the real-time performance is poor can be effectively solved.

Description

Technical Field

[0001] The invention belongs to the technical field of fire prediction, and in particular relates to a fire spread prediction method based on time-space point process. Background Art

[0002] In recent years, forest fires have occurred frequently around the world, posing a huge threat to ecosystems and human life. Therefore, how to accurately predict the dynamics of fire spread and take timely preventive and firefighting measures has become the focus of research and practice in various countries. Existing fire prediction technologies can be roughly divided into three categories: physical models, semi-empirical models, and empirical models. These models each have different advantages and application scenarios, but they still have many limitations when facing complex and dynamic fire spread.

[0003] Physical models, such as the WRF-FIRE model and the Coupled Atmosphere-Wildland Fire Environment system, are based on the physical and chemical principles of the fire combustion process and can describe the physical phenomena in the process of fire spread in detail. However, these models rely on a large amount of environmental data (such as meteorological data, topographic data, etc.), the model calculation is complex, the scope of application is limited, and the real-time performance is poor. Semi-empirical models combine physical processes and observational data to predict fire conditions by establishing a functional relationship of fire spread. Although these models, such as the Rothermel fire spread model, are relatively simple and have a small amount of calculation, they rely on empirical parameters and cannot adapt to changing environmental conditions. In addition, empirical models are mainly based on statistical analysis of historical data and use the relationship between historical fire data and geographical and meteorological conditions to predict fire conditions. Although these models are computationally efficient, they are difficult to generalize to different ecosystems or environments and have poor universality.

[0004] Although existing technologies have made significant progress in fire prediction, they still face the following problems and challenges: Insufficient capture of local and global spatiotemporal dependencies: Fire spread is not only affected by the local environment (such as fuel distribution and wind speed), but also by remote factors (such as spark ignition). Existing models often have difficulty capturing local and remote spatiotemporal correlations at the same time, especially when faced with complex terrain and meteorological conditions.

[0005] Poor generalization of the model: Although the physical model can accurately simulate the spread of fire under specific conditions, its ability to predict large-scale and diverse fires is limited. The prediction accuracy of empirical and semi-empirical models is often limited by the regionality of the data, making it difficult to promote and apply them under different environmental conditions.

[0006] High computational cost and poor real-time performance: Physical models and semi-physical models often require a large amount of computing resources when calculating complex fire spread processes, making it difficult to provide real-time fire predictions in emergency situations.

[0007] In response to these problems, the challenges faced by current research are mainly reflected in how to capture the spatiotemporal characteristics of local and remote fire spread at a low computational cost, and how to provide accurate fire prediction under diverse ecological and environmental conditions. Although machine learning and deep learning technologies have achieved certain results in fire prediction, existing models based on convolutional neural networks (CNNs) and traditional LSTMs still have difficulty in fully describing the spatiotemporal dynamics of fires. Summary of the invention

[0008] In order to solve the above problems, the present invention proposes a fire spread prediction method based on spatiotemporal point processes, which captures local fire dynamics through a convolutional long short-term memory network and processes long-range spatiotemporal correlations through a self-attention mechanism, thereby improving the prediction accuracy of fire spread. This method can effectively solve the problems of insufficient capture of local and global spatiotemporal dependencies, poor model generalization ability, high computational cost and poor real-time performance in the prior art.

[0009] To achieve the above object, the technical solution adopted by the present invention is: A fire spread prediction method based on time-space point process, comprising the steps of: By observing each time point, we extract multidimensional features and embed them into the ConvLSTM network to form spatiotemporal series data of fire spread. The local spatiotemporal features are extracted through the ConvLSTM network to capture local dependencies. Introducing the self-attention mechanism, calculating the dependencies between grids, interacting with fire conditions in remote areas, and capturing remote dependencies; After being processed by the ConvLSTM network and the self-attention mechanism, the output is the prediction result of the fire situation at the future moment.

[0010] Furthermore, in the ConvLSTM network, the spatial features and temporal features of each grid cell in the fire spreading process are combined through convolution operation to achieve spatiotemporal dependency modeling of the fire, including: Input layer: The input includes the fire status and environmental characteristics of the past few days; ConvLSTM unit: Each ConvLSTM unit contains an input gate, a forget gate, an output gate, and a cell state connected in sequence, where the convolution operation is used to extract spatial features and the LSTM unit is used to process time series information.

[0011] Furthermore, the self-attention mechanism is used to emphasize the influence of remote areas on the current fire status by calculating the dependency matrix between grids; the self-attention mechanism adopts point-to-point self-attention, including: By mapping the input features into three feature spaces of query, key and value, the similarity scores between the input grid cells are calculated to generate a weight matrix, which is used to weight the input features; Aiming at the spatiotemporal dependency of local space, a local region-based attention mechanism is established. The grid is divided into several local regions and the attention weights within the region are calculated to capture the local fine-grained dependency.

[0012] Furthermore, after being processed by the ConvLSTM network and the self-attention mechanism, the output is the prediction result of the fire situation at the future moment: the output layer uses 3D convolution operation to aggregate the spatiotemporal features, generate the probability distribution of the fire situation at the next moment, and predict the spread direction and position of the fire front.

[0013] Furthermore, the output layer outputs a two-dimensional grid with the same spatial dimension as the input, where each cell in the grid represents the probability of fire spread at a certain time step in the future; through a preset threshold, the prediction result is converted into a classification result of the fire spread area.

[0014] The beneficial effects of adopting this technical solution are: By combining the convolutional long short-term memory network (ConvLSTM) with the self-attention mechanism, the present invention can simultaneously capture the local dynamics of fire spread and remote ignition phenomena, and can effectively cope with the spatiotemporal complexity of the fire spread process, thereby improving the accuracy of fire spread prediction. While ensuring high-precision prediction, the present invention has good real-time and adaptability, and is suitable for early warning and monitoring of various fire scenarios.

[0015] The present invention can effectively solve the problem of insufficient capture of local and remote dependencies in the prior art by combining ConvLSTM and point-to-point self-attention mechanism. ConvLSTM extracts spatial features through convolution operations, and combined with the time series processing capabilities of LSTM, it can effectively deal with the spread of the fire front; and the self-attention mechanism can accurately handle remote ignition phenomena through dynamic weighting. The combination of these technical means enables the fire prediction model to capture local fire dynamics and consider the interaction of remote fires, thereby significantly improving the accuracy and adaptability of the prediction. Therefore, the present invention has significant advantages at the social, economic and technical levels. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The present invention is a flow chart of a method for predicting fire spread based on a time-space point process. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described below with reference to the accompanying drawings.

[0018] In this embodiment, see Figure 1 As shown, the present invention proposes a fire spread prediction method based on a spatiotemporal point process, comprising the steps of: By observing each time point, multidimensional features (such as fuel distribution, wind speed, humidity, terrain, etc.) are extracted and embedded into the ConvLSTM network to form spatiotemporal series data of fire spread; Integration of multidimensional environmental features: By combining multiple environmental features, such as fuel status, wind speed, humidity, etc., the model's ability to predict complex fire dynamics is enhanced, and its generalization performance in different ecosystems is improved.

[0019] The local spatiotemporal features are extracted through the ConvLSTM network to capture local dependencies. ConvLSTM combines the ability of convolutional neural networks (CNN) to extract spatial information and the ability of LSTM to remember time series data, and can effectively capture the spread path of the fire front in the local grid. It is used to extract the local spatiotemporal dependency features of fire spread and solve the problem of the gradual spread of fire in neighboring areas.

[0020] Since the spark ignition phenomenon may occur in an area far away from the fire front, the present invention introduces a self-attention mechanism, calculates the dependencies between grids, conducts fire interaction in remote areas, and realizes the capture of remote dependencies. The self-attention mechanism is used to highlight the importance of remote ignition events in the spread of fire through a dynamic weighted approach. The method is used to capture remote dependencies, especially the long-distance spread effect caused by the spark ignition phenomenon, to address the deficiency that the prior art cannot handle local and remote dependencies at the same time.

[0021] After being processed by the ConvLSTM network and the self-attention mechanism, the output is the prediction result of the fire situation at the future moment. Specific embodiment:

[0023] Data sources: Fire data can come from satellite remote sensing observations (such as VIIRS, MODIS, etc.) or historical fire records, while environmental data include meteorological data (wind speed, temperature, humidity, etc.) and terrain data (slope, vegetation type, etc.). All data must ensure temporal and spatial resolution and integrity.

[0024] Data format: Organize the fire data in a spatial grid format, such as dividing the observation area into a two-dimensional grid of X×YX \timesYX×Y. The size of the grid unit can be adjusted according to the specific application scenario (such as 10 meters, 100 meters, 1 kilometer). The time dimension is defined in days, hours, or minutes, depending on the required time series accuracy.

[0025] Feature extraction: Fuel distribution: Extract fuel characteristics for each grid based on vegetation density and type data. Wind speed and direction: Extract wind speed and direction information for each grid from meteorological data as key drivers of fire spread. Humidity: Extract humidity information from meteorological data as an important variable affecting fire spread. Terrain: Extract slope and elevation information from terrain data as auxiliary features affecting fire spread.

[0026] As an optimization solution of the above embodiment, the ConvLSTM network is used to capture the local spatiotemporal characteristics of the fire spread. The spatial characteristics and temporal characteristics of each grid cell in the fire spread process are combined through convolution operation to achieve spatiotemporal dependency modeling of the fire, including: Input layer: The input includes the fire status and environmental characteristics of the past few days. Each input sample is a tensor of X×Y×TX \times Y \times TX×Y×T, where XXX and YYY represent the number of rows and columns of the spatial grid, and TTT represents the time step. The input data can be multidimensional data such as fuel, wind speed, humidity, etc.

[0027] ConvLSTM unit: Each ConvLSTM unit contains an input gate, a forget gate, an output gate, and a cell state connected in sequence, where the convolution operation is used to extract spatial features and the LSTM unit is used to process time series information. This structure can capture the spatiotemporal dependencies between local grids.

[0028] Each ConvLSTM unit extracts spatial features through convolution operations and processes time series data through LSTM units.

[0029] The convolution kernel size is 3×33 \times 33×3 or 5×55 \times 55×5 (depending on the resolution of the fire area).

[0030] The time step number TTT of the ConvLSTM layer is 5 to 10 days, depending on the needs of fire spread prediction.

[0031] As an optimization solution of the above embodiment, in order to solve the problem that the existing model is difficult to capture remote dependencies, the present invention introduces the use of a self-attention mechanism to emphasize the impact of remote areas on the current fire status by calculating the dependency matrix between grids; the self-attention mechanism uses point-to-point self-attention to capture the spatiotemporal dependencies between remote areas, including: By mapping the features output by the ConvLSTM layer as input features into three feature spaces: query, key, and value, the similarity scores between the input grid cells are calculated, and a weight matrix is ​​generated. This matrix is ​​used to weight the input features, so that the impact of the fire in the remote area on the current grid can be better captured. Aiming at the spatiotemporal dependency of local space, a local area-based attention mechanism is established to improve the accuracy of local fire spread. The grid is divided into several local areas, and the attention weights within the area are calculated to capture the local fine-grained dependency.

[0032] The size of each local region can be set to a 7×77 \times 77×7 or 9×99 \times 99×9 grid for more refined handling of local spreading.

[0033] As an optimization scheme of the above embodiment, after being processed by the ConvLSTM network and the self-attention mechanism, the output is the prediction result of the fire situation at the future moment: the output layer uses a 3D convolution operation to aggregate the spatiotemporal features, generate the probability distribution of the fire situation at the next moment, and predict the spread direction and position of the fire front.

[0034] Output layer output format: The output is a two-dimensional grid with the same spatial dimension as the input. Each cell in the grid represents the probability value of fire spread at a certain time step in the future. Through the preset threshold, the prediction result is converted into the classification result of the fire spread area.

[0035] In summary, the model structure includes: ConvLSTM layer, self-attention layer and output layer.

[0036] Model training: Training data set: Use a known fire data set for model training. Each sample is a three-dimensional tensor of X×Y×TX \times Y \times TX×Y×T, which contains spatial and temporal dimensions. The label of each sample is the fire spread state (fire and non-fire area) at the corresponding time step.

[0037] Loss function: Use mean square error (MSE) or binary cross-entropy as the loss function to measure the deviation between the predicted results and the actual fire data.

[0038] Optimizer: The Adam optimizer is used with the learning rate set to 0.001 and gradually decreased to 10−510^{-5}10−5 to speed up the training process and prevent overfitting.

[0039] Training steps: Each training is iterated using samples with a batch size of 32 to 64. The number of steps for model training is determined by the size of the dataset and the prediction accuracy requirements, usually 100 to 300 epochs.

[0040] Prediction and Application: Prediction process: The pre-processed historical fire data is input into the trained model, and the model will predict the probability of fire spread in the next few days based on the input data.

[0041] Prediction result format: The output is a two-dimensional fire probability map with the same spatial size as the input. The value of each grid cell is a probability value between 0 and 1, indicating the probability of a fire occurring in the area at some point in the future.

[0042] Threshold setting: In order to determine whether a fire has occurred, a probability threshold (such as 0.5) can be set. If the predicted probability of a grid cell is greater than the threshold, it is judged that a fire may have occurred in the area.

[0043] Model evaluation: Evaluate model performance on real fire data sets, using indicators such as F1 score, precision, and recall to measure the model's prediction effect. Optimize prediction results by comparing model performance at different thresholds and in different regions.

[0044] Compared with traditional physical models and empirical models, the present invention has significant advantages in dealing with complex spatiotemporal dependencies, which are specifically manifested in the following aspects: (1) Higher prediction accuracy: The present invention captures the local spatiotemporal dependency of the fire through ConvLSTM and effectively handles the long-range dependency through the self-attention mechanism, especially when there is a spark ignition phenomenon, showing a better prediction effect. According to experimental results, the method of the present invention can reduce the mean mean square error (RMSE) of fire prediction by about 10%-15%, and improve the prediction accuracy of the fire front to more than 90%.

[0045] (2) Adaptability to complex environments: The present invention demonstrates good generalization capabilities in different ecological environments. Compared with traditional models, the generalization performance of the present invention is improved by about 20%, and it can effectively predict the spread of fire in different geographical regions and climatic conditions, and is applicable to a variety of fire scenarios.

[0046] (3) Reduce computational costs: Traditional physical models have high computational complexity when simulating large-scale fire spread, making it difficult to achieve real-time prediction. This invention is based on deep learning technology. Although the amount of computation is large during the training phase, the computational complexity is significantly reduced during the actual prediction phase, which can achieve rapid fire prediction and meet real-time monitoring needs.

[0047] By improving the accuracy and real-time performance of fire prediction, the present invention can effectively reduce the economic losses caused by fire: (1) Improved fire warning efficiency: Since the spread of fire can be predicted more accurately, the present invention helps to issue early warnings in the early stages of a fire outbreak, thereby reducing property losses and firefighting costs. According to relevant statistics, early warning can reduce firefighting costs by about 25%. The improvement of fire prediction accuracy can also reduce the waste of resources caused by inaccurate predictions.

[0048] (2) Reduce fire prevention costs: The present invention provides a more intelligent fire prediction system to help relevant departments optimize fire fighting strategies, rationally allocate resources, and reduce unnecessary resource investment. It is estimated that through effective prediction and resource scheduling, the cost of fire prevention can be reduced by about 15%-20%.

[0049] Social effect: The present invention can provide accurate fire prediction in various fire scenarios, effectively reducing casualties and ecological damage caused by forest fires: (1) Reducing casualties: Through more accurate prediction of fire spread, the present invention can help fire prevention departments to evacuate residents in disaster areas in a timely manner, reducing the risk of casualties. According to experimental results, timely warning after a fire breaks out can advance the evacuation time of people in the disaster area by about 2 hours, effectively reducing potential safety threats.

[0050] (2) Protecting ecosystems: Forest fires are extremely destructive to the ecological environment. The present invention can help formulate timely protection measures to avoid the destruction of important ecosystems by accurately predicting the spread of fire. It is estimated that reasonable fire prevention measures can reduce vegetation loss by about 30%, which helps to protect carbon sinks and biodiversity.

[0051] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A fire spread prediction method based on spatiotemporal point process, characterized in that: Includes steps: By observing each time point, we extract multidimensional features and embed them into the ConvLSTM network to form spatiotemporal series data of fire spread. The local spatiotemporal features are extracted through the ConvLSTM network to capture local dependencies. Introducing the self-attention mechanism, calculating the dependencies between grids, interacting with fire conditions in remote areas, and capturing remote dependencies; After being processed by the ConvLSTM network and the self-attention mechanism, the output is the prediction result of the fire situation in the future.

2. A fire spread prediction method based on spatiotemporal point process according to claim 1, characterized in that: In the ConvLSTM network, the spatial features and temporal features of each grid cell in the fire spreading process are combined through convolution operations to achieve spatiotemporal dependency modeling of the fire situation, including: Input layer: The input includes the fire status and environmental characteristics of the past few days; ConvLSTM unit: Each ConvLSTM unit contains an input gate, a forget gate, an output gate, and a cell state connected in sequence, where the convolution operation is used to extract spatial features and the LSTM unit is used to process time series information.

3. The fire spread prediction method based on spatiotemporal point process according to claim 1 is characterized in that: Using the self-attention mechanism, the influence of remote areas on the current fire status is emphasized by calculating the dependency matrix between grids; The self-attention mechanism uses point-to-point self-attention, including: By mapping the input features into three feature spaces of query, key and value, the similarity scores between the input grid cells are calculated to generate a weight matrix, which is used to weight the input features; Aiming at the spatiotemporal dependency of local space, a local region-based attention mechanism is established. The grid is divided into several local regions and the attention weights within the region are calculated to capture the local fine-grained dependency.

4. The fire spread prediction method based on spatiotemporal point process according to claim 1, characterized in that: After being processed by the ConvLSTM network and the self-attention mechanism, the output is the prediction result of the fire situation at the future moment: the output layer uses 3D convolution operation to aggregate the spatiotemporal features, generate the probability distribution of the fire situation at the next moment, and predict the spread direction and position of the fire front.

5. The fire spread prediction method based on spatiotemporal point process according to claim 4 is characterized in that: Output layer output format: The output is a two-dimensional grid with the same spatial dimension as the input, and each cell in the grid represents the probability value of fire spread at a certain time step in the future; By setting a preset threshold, the prediction results are converted into classification results of the fire spread area.

Images