Spatial distribution prediction method of urban fire incidence based on deep learning algorithm

Abstract

The present invention discloses a method for predicting the spatial distribution of urban fires based on a deep learning algorithm, comprising the following steps: collecting raw data; preprocessing the raw data; determining input variables and output variables for an urban fire forecast model, wherein the input variables include urban land type raster data, meteorological factor time series data for the past several days, and a set of daily fire count raster maps for the past several days; and output variables are a raster map of the cumulative fire count for the next several days. Furthermore, the method constructs an urban fire forecast model based on a deep neural network, wherein the urban fire forecast model includes an encoder for processing time series variables and a decoder for processing urban land type spatial variables and feature variables. Furthermore, the method trains and evaluates the urban fire forecast model. The present invention can jointly process temporal and spatial features, achieves high prediction accuracy, and effectively improves the accuracy of urban fire risk prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of urban fire prediction, and in particular to a method for predicting the spatial distribution of urban fire numbers based on a deep learning algorithm. Background Art

[0002] With the acceleration of urbanization, urban fire risk is becoming increasingly significant. Facing this challenge, predicting potential fire risks in urban areas can support urban fire management and fire emergency decision-making. For example, it can provide a scientific basis for measures such as the pre-positioning of firefighting resources, thereby reducing the probability of fire and minimizing fire losses. However, current research on urban fire risk has largely focused on calculating fire risk indices for individual buildings, with limited efforts to develop regional-scale fire spatial distribution prediction models from the perspective of the entire city.

[0003] Compared to wildfires, urban fires exhibit more complex dynamics and lower predictability due to the influence of more factors. The factors influencing urban fire distribution fall into two main categories: spatial and temporal. Spatially, factors such as a city's population distribution, building types, and functional zoning can all influence urban fire risk. For example, urban land use classifications include residential, commercial, industrial, and green spaces, and the triggers for fires vary across land use types. Temporally, fire incidence exhibits cyclical variations with seasonal climate changes. Historical fire records show that the probability of fires is higher in dry summer weather than in other seasons. Furthermore, the probability of fires occurring during the day is often higher than at night. In summary, urban fires have numerous types and complex mechanisms. Using known fire-influencing factors to infer the spatial distribution of urban fires over the future presents a significant challenge. Although some data-driven methods have been applied to urban fire prediction tasks, most of them only consider temporal features or only spatial features, and only a few methods consider both temporal and spatial features. However, even these methods still have shortcomings in the joint modeling of spatial and temporal features, resulting in inaccurate performance in fire prediction.

[0004] Artificial intelligence technologies such as machine learning are expected to overcome the shortcomings of traditional methods in capturing complex disaster mechanisms, establish a mapping relationship between key fire influencing factors and the spatial distribution of fire occurrences, and provide technical support for improving the accuracy of urban fire predictions. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of traditional technologies in predicting and accurately capturing the spatial distribution patterns of urban fires in the future, and to provide a method for predicting the spatial distribution of urban fires based on a deep learning algorithm.

[0006] The technical solution adopted by the present invention to overcome the technical problems is:

[0007] A method for predicting the spatial distribution of urban fire incidence based on a deep learning algorithm includes the following steps:

[0008] S1. Collect original data: Collect historical fire event data in the urban study area and urban land use type data and meteorological factor data corresponding to each historical fire event;

[0009] S2. Preprocessing the raw data collected in step S1: Divide the urban study area into grid cells and perform spatiotemporal aggregation on the historical fire event data in the urban study area; Calculate the area ratio of each land use type in each grid cell; Average the meteorological factor data recorded by all meteorological stations in the urban study area;

[0010] S3, based on the data pre-processed in step S2, determine the input variables and output variables for the urban fire number forecast model: the input variables include urban land type raster data, past Time series data of meteorological factors of the day and past The three types of input variables are spatial type data, time type data and spatiotemporal type data; the output variable is the future Daily cumulative fire number raster map; , ,and and All are integers;

[0011] S4. Constructing a city fire number prediction model based on a deep neural network: The city fire number prediction model includes an encoder and a decoder, wherein the encoder is used to process time series variables, and the decoder is used to process urban land type spatial variables and feature variables;

[0012] S5. Training and evaluation of the urban fire number forecast model: Arrange the input variables and output variables of step S3 to obtain paired data sets, and divide the data sets into a training set, a validation set, and a test set in chronological order and at a preset ratio. The training set and validation set are used to determine the parameters of the urban fire number forecast model, and the test set is used to evaluate the performance of the urban fire number forecast model.

[0013] Furthermore, in step S1, the historical fire event data of the urban study area includes the address and time of occurrence of each fire; the urban land type data includes two categories: urban and rural land and urban construction land, among which urban and rural land is subdivided into construction land and non-construction land, and urban construction land is subdivided into residential land, public administration and public service land, commercial service facility land, industrial land, logistics and warehousing land, road and transportation facility land, public utility land, green space and square land according to function; the meteorological factor data includes the daily average temperature, daily average relative humidity and daily cumulative rainfall recorded by all meteorological observation stations in the urban study area.

[0014] Furthermore, in step S2, the urban study area is divided into grid cells, and spatiotemporal aggregation processing is performed on the historical fire event data of the urban study area, specifically including:

[0015] In terms of time, the data is divided into daily statistics to form a daily fire count data set;

[0016] In space, the grid cells are divided according to the preset spatial resolution, and the geographical coordinate points of the daily fire count are counted into the grid cells to form a spatial grid map of the daily fire count.

[0017] Furthermore, in step S2, for meteorological factors, the meteorological factor data recorded by all meteorological stations in the urban study area are averaged, specifically including:

[0018] For different meteorological factors, the data recorded by all meteorological stations in the urban study area were averaged to obtain the time series data of daily average temperature, daily average relative humidity and daily cumulative rainfall in the urban study area.

[0019] Furthermore, in step S3, the urban land type raster data is as follows:

[0020] The urban land type raster maps are stacked into multi-channel image data in the spatial dimension. Assume that the set of all urban land type raster layers is represented as follows:

[0021] (1)

[0022] In formula (1), Represents one of the urban land use type layers, , let the length and width of each raster layer be and , then the variable Dimension size ,Right now, As the input variable of the urban fire number prediction model.

[0023] Further, in step S3, the past The time series data of meteorological factors for the day are as follows:

[0024] The meteorological factors are stacked into multi-feature time series data in the time dimension. Let the time of the day be ,past The time series data of meteorological factors for a day are expressed as follows:

[0025] (2)

[0026] In formula (2), Indicates the Time series of meteorological factors of the day, , the meteorological factor time series specifically includes three variables: daily average temperature, daily average relative humidity and daily cumulative rainfall, so the variable The dimension size is ,Right now, As the input variable of the urban fire number prediction model.

[0027] Further, in step S3, the past The daily fire count raster map set is as follows:

[0028] past The daily fire count raster map set is represented as follows:

[0029] (3)

[0030] In formula (3), Indicates the Daily fire count raster map, ;

[0031] The empirical orthogonal function decomposition method is used to transform the spatiotemporal data into Perform principal component decomposition and retain the previous principal components, so that the The cumulative variance explanation rate of the principal components is not less than the preset percentage, as follows:

[0032] (4)

[0033] In formula (4), Indicates the number of grids in the daily fire count grid map, Represents the spatial component of the decomposed daily fire count raster map set, Represents the time component of the decomposed daily fire count raster map set, extract As the input variable of the urban fire number prediction model.

[0034] Furthermore, in step S3, in the future The daily cumulative fire count raster map is as follows:

[0035] By moving the present day into the future Total days The daily fire count grid map collection is aggregated in time to form the future The daily cumulative number of fires grid map is:

[0036] (5)

[0037] In formula (5), Indicates the The daily fire count raster map will be used to The daily cumulative fire number raster map is used as the output variable of the urban fire number prediction model.

[0038] Furthermore, in step S4, the architecture of the encoder includes a long short-term memory network connected to a self-attention mechanism network, and the architecture of the decoder includes a fully connected network and several layers of convolutional neural networks;

[0039] The time series components that the encoder processes include meteorological factor time variables and the time component of the historical daily fire count Specific operations include:

[0040] The meteorological factor time variable and the time component of the historical daily fire count raster map collection Splicing in the time dimension to obtain a composite time series [ E, M ] , composite time series [ E, M ] The dimension is , composite time series [ E, M ] As the input variable of the encoder, the encoder processes the composite time series [ E, M ] Encode, and then use the last moment variable of the self-attention mechanism network output as the feature variable , where r is the number of hidden units in the self-attention mechanism network;

[0041] Decoder used to process urban land use type spatial variables and characteristic variables Specific operations include:

[0042] First, the characteristic variables For preprocessing, a fully connected network is used to Transformed into Secondly, use matrix transformation to Further transformed into variables ; Then, the variable and variables Splicing on the channel dimension to obtain composite spatial variables [ Z ̃ , L] , composite spatial variables [ Z ̃ , L] The dimension is ; Input the composite spatial variable into the multi-layer convolutional neural network, and the output of the last layer of convolutional neural network is the future The predicted value of the daily cumulative fire number grid map is set as .

[0043] Furthermore, in step S5, the training of the urban fire number prediction model specifically includes:

[0044] Assume that the loss function of the urban fire number forecast model is the mean square error (MSE) of the district-level disaster loss, and the number of training samples is N, where A sample of the future The actual and predicted values of the daily cumulative fire number grid map are and , then the loss function is expressed as follows:

[0045] (6)

[0046] Set a number of training iterations, use the training set to backpropagate the mean square error (MSE) of the urban fire occurrence forecast model using the optimizer, and use the validation set to calculate the loss function of the urban fire occurrence forecast model after each round of training. When the number of training iterations of the urban fire occurrence forecast model is reached, stop training; otherwise, continue training the urban fire occurrence forecast model. Within the iteration cycle, select the urban fire occurrence forecast model with the best prediction result on the validation set as the final trained urban fire occurrence forecast model, and determine the parameters of the urban fire occurrence forecast model.

[0047] The beneficial effects of the present invention are:

[0048] The present invention can effectively capture the spatial distribution characteristics of the expected number of fires at a certain spatial resolution within a certain period of time in the future. The present invention fully considers the spatiotemporal distribution of urban land use types, meteorological factors, and the number of daily fires in the past few days as predictive variables of the urban fire number forecast model, covering the main factors affecting the occurrence of urban fires. In addition, the present invention proposes an urban fire number forecast model based on a deep neural network, including an encoder and a decoder, which can jointly process temporal features and spatial features, significantly improving the adaptability and prediction accuracy of the forecast model to different types of data, and effectively improving the accuracy of urban fire risk forecasting. In addition, predicting the potential number of fires in a city in the future and their spatial distribution is of great significance to urban fire prevention and control and fire emergency management. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flowchart of the method for predicting the spatial distribution of urban fire numbers based on a deep learning algorithm according to an embodiment of the present invention.

[0050] Figure 2 This is a schematic diagram of converting original historical fire event point data into a daily fire count spatial grid map after preprocessing in an embodiment of the present invention.

[0051] Figure 3 Schematic diagram of the architecture of the urban fire number forecast model described in an embodiment of the present invention.

[0052] Figure 4 This is the comparison result between the actual value and the predicted value of the spatial distribution of the cumulative number of fires in the next 30 days in the embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to facilitate those skilled in the art to better understand the present invention, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. The following is only exemplary and does not limit the scope of protection of the present invention.

[0054] The present invention discloses a method for predicting the spatial distribution of urban fire incidence based on a deep learning algorithm, comprising the following steps:

[0055] S1. Collect original data: Collect historical fire event data in the urban study area and urban land use type data and meteorological factor data corresponding to each historical fire event;

[0056] S2. Preprocessing the raw data collected in step S1: Divide the urban study area into grid cells and perform spatiotemporal aggregation on the historical fire event data in the urban study area; Calculate the area ratio of each land use type in each grid cell; Average the meteorological factor data recorded by all meteorological stations in the urban study area;

[0057] S3, based on the data pre-processed in step S2, determine the input variables and output variables for the urban fire number forecast model: the input variables include urban land type raster data, past Time series data of meteorological factors of the day and past The three types of input variables are spatial type data, time type data and spatiotemporal type data; the output variable is the future Daily cumulative fire number raster map; , ,and and All are integers;

[0058] S4. Constructing a city fire number prediction model based on a deep neural network: The city fire number prediction model includes an encoder and a decoder, wherein the encoder is used to process time series variables, and the decoder is used to process urban land type spatial variables and feature variables;

[0059] S5. Training and evaluation of the urban fire number forecast model: Arrange the input variables and output variables of step S3 to obtain paired data sets, and divide the data sets into a training set, a validation set, and a test set in chronological order and at a preset ratio. The training set and validation set are used to determine the parameters of the urban fire number forecast model, and the test set is used to evaluate the performance of the urban fire number forecast model.

[0060] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in detail below with reference to the accompanying drawings. These are merely exemplary embodiments of the present invention. However, it should be understood that the present invention can be implemented in various forms and is not limited to the embodiments described herein. These embodiments are intended to enable those skilled in the art to understand the present invention more clearly and thoroughly.

[0061] This embodiment describes a method for predicting the spatial distribution of urban fires based on a deep learning algorithm. Figure 1 As shown, the following steps are included:

[0062] S1. Collect original data: Collect historical fire event data in the urban study area and urban land use type data and meteorological factor data corresponding to each historical fire event.

[0063] In this embodiment, historical fire event data that occurred in a certain urban study area over the past 10 years were collected, including the location and time of occurrence of each fire.

[0064] Vector data of land use types in the urban study area were collected, including construction land, non-construction land, residential land, public administration and public service land, commercial service facility land, industrial land, logistics and warehousing land, road and transportation facility land, public utility land, green space and square land, a total of 10 land use types.

[0065] Daily meteorological observation datasets were obtained through official channels, such as the official National Oceanic and Atmospheric Administration, and meteorological factors recorded by meteorological stations in the urban study area were extracted, including daily average temperature, daily average relative humidity, and daily accumulated rainfall.

[0066] S2. Preprocess the raw data collected in step S1: divide the urban study area into grid cells, and perform spatiotemporal aggregation processing on the historical fire event data in the urban study area; for urban land use types, calculate the area ratio of each type of land use in each grid; for meteorological factors, average the meteorological factor data recorded by all meteorological stations in the urban study area.

[0067] (1) The urban study area is divided into grid cells, and the historical fire event data of the urban study area are subjected to spatiotemporal aggregation processing, including:

[0068] If the occurrence address of the historical fire event data collected in step S1 is a geographic coordinate, it can be used directly; otherwise, a map tool is used to convert the occurrence address attributes of the historical fire event data into geographic coordinates.

[0069] In terms of time, the data is divided into daily statistics to form a daily fire count data set.

[0070] In space, the grid cells are divided according to the preset spatial resolution, and the geographical coordinate points of the daily fire count are counted into the grid cells to form a daily fire count spatial grid map. In this embodiment, the preset spatial resolution is preferably 2.5 kilometers, and the length of each grid layer is 1.5 kilometers. He Kuan Both are 20, that is, each grid layer has 20 grids in both length and width directions, forming a spatial grid map of the number of daily fires with an image size of 20×20, such as Figure 2 As shown, Figure 2 The left figure shows the raw point data of the number of fires per day. Figure 2 The right figure shows the spatial distribution data of the daily fire count on the grid.

[0071] (2) For urban land use types, calculate the area ratio of each type of land use in each grid, including:

[0072] For urban land types, the geographic spatial analysis tool ArcGIS was used to calculate the area ratio of each urban land type in each grid.

[0073] (3) For meteorological factors, the meteorological factor data recorded by all meteorological stations in the urban study area are averaged, including:

[0074] For different meteorological factors, the data recorded by all meteorological stations within the urban study area are averaged to obtain time series data for the daily average temperature, daily average relative humidity, and daily cumulative rainfall of the urban study area. Since only one meteorological station in the urban study area is collected in this example, this step is skipped. Only the meteorological factor data recorded by this one meteorological station in the urban study area need to be collected. In other words, only the time series data for the daily average temperature, daily average relative humidity, and daily cumulative rainfall of this meteorological station are collected.

[0075] S3, based on the data pre-processed in step S2, determine the input variables and output variables for the urban fire number forecast model: the input variables include urban land type raster data, past Time series data of meteorological factors of the day and past The three types of input variables are spatial type data, time type data and spatiotemporal type data; the output variable is the future Daily cumulative fire number raster map; The value range is ; Is a hyperparameter, and the optimal value can be found through sensitivity analysis. Here The minimum value of is 1 day, that is ,and and All are integers.

[0076] In this embodiment, it is preferred , .

[0077] (1) For urban land use types, the urban land use type raster data are as follows:

[0078] In this embodiment, there are 10 types of urban land use, so the urban land use type raster map has 10 layers. The 10 layers of urban land use type raster maps are stacked into 10-channel image data in the spatial dimension. Suppose that the set of all urban land use type raster layers is represented as follows:

[0079] (1)

[0080] In formula (1), Represents one of the urban land use type layers, , at this time the variable The dimension size is 10×20×20, that is, As the input variable of the urban fire number prediction model.

[0081] (2) The meteorological factor time series of the past 30 days, namely, daily average temperature, daily average relative humidity, and daily cumulative rainfall, are spliced into three feature time series in the time dimension. The data are as follows:

[0082] Assume that the time of the day is , the time series data of meteorological factors in the past 30 days are expressed as follows:

[0083] (2)

[0084] In formula (2), Indicates the Time series of meteorological factors of the day, , the meteorological factor time series specifically includes three variables: daily average temperature, daily average relative humidity and daily cumulative rainfall, so the variable The dimension size is 3×30, that is, As the input variable of the urban fire number prediction model.

[0085] (3) The daily fire count grid map set for the past 30 days is as follows:

[0086] The daily fire count raster map collection for the past 30 days is shown below:

[0087] (3)

[0088] In formula (3), Indicates the Daily fire count raster map, ;

[0089] The empirical orthogonal function decomposition method is used to transform the spatiotemporal data into Perform principal component decomposition and retain the previous principal components, so that the The cumulative variance explanation rate of the principal components is not less than a preset percentage, which is generally at least 95%. The specific number of principal components retained depends on the data situation. In this embodiment, the cumulative variance explanation rate of the principal components retained is preferably 98.5%. Then, when k=100, the cumulative variance explanation rate of the first 100 principal components is 98.5%, as follows:

[0090] (4)

[0091] In formula (4), Represents the number of grids in the daily fire count grid map, i.e. m=400, Represents the spatial component of the decomposed daily fire count raster map set, Represents the time component of the decomposed daily fire count raster map set, extract As the input variable of the urban fire number forecast model, it is finally decomposed into the spatial components of the daily fire number raster map set. The time component of the daily fire count raster map collection .Will As the input variable of the urban fire number prediction model.

[0092] (4) The grid map of the cumulative number of fires in the next 30 days is as follows:

[0093] By temporally aggregating the daily fire count grid graphs from the current day to the next 29 days, a cumulative fire count grid graph for the next 30 days is formed, namely:

[0094] (5)

[0095] In formula (5), Indicates the The daily fire number grid map for the next 30 days is used as the output variable of the urban fire number forecast model.

[0096] S4. Construct a city fire number prediction model based on deep neural network: The city fire number prediction model includes an encoder and a decoder. The encoder is used to process time series variables, and the decoder is used to process urban land type spatial variables and feature variables.

[0097] In this embodiment, Figure 3 As shown in the figure, the architecture of the encoder includes a long short-term memory network connected to a self-attention mechanism network, and the architecture of the decoder includes a fully connected network and several layers of convolutional neural networks.

[0098] (1) The time series components processed by the encoder include meteorological factor time variables and the time component of the historical daily fire count Specific operations include:

[0099] The meteorological factor time variable and the time component of the historical daily fire count raster map collection Splicing in the time dimension to obtain a composite time series [ E, M ] , composite time series [ E, M ] The dimension is 103×30, a composite time series [ E, M ] As the input variable of the encoder, the encoder processes the composite time series [ E, M ] Encode, and then use the last moment variable of the self-attention mechanism network output as the feature variable , where r is the number of hidden units in the self-attention network. In this embodiment, the hidden units of the first-layer long short-term memory network and the second-layer self-attention network are both 256. The tensor of size 103×30 is transformed into a tensor of size 256×30. The variable at the last moment of the attention network output is taken as the output of the encoder, which is recorded as .

[0100] (2) Decoder is used to process urban land type spatial variables and characteristic variables Specific operations include:

[0101] First, the characteristic variables For preprocessing, a fully connected network is used to Transformed into Secondly, use matrix transformation to Further transformed into variables ; Then, the variable and variables Splicing on the channel dimension to obtain composite spatial variables [ Z ̃ , L] , composite spatial variables [ Z ̃ , L] The dimension is 11×20×20; the decoder of this embodiment uses a two-layer convolutional neural network to input the composite spatial variable into the two-layer convolutional neural network. The first layer of the convolutional neural network has 64 convolution kernels of size 3×3, and the second layer of the convolutional neural network has 1 convolution kernel of size 2×2. The two-layer convolution operation converts the original tensor of size 11×20×20 into a tensor of size 1×20×20, and the generated grid map of the cumulative number of fires in the next 30 days is predicted. as the output of the decoder.

[0102] S5. Training and evaluation of the urban fire number forecast model: Arrange the input variables and output variables of step S3 to obtain paired data sets, and divide the data sets into training set, validation set and test set in chronological order and at a preset ratio of 8:1:1. The training set and validation set are used to determine the parameters of the urban fire number forecast model, and the test set is used to evaluate the performance of the urban fire number forecast model.

[0103] Specifically, the training of the urban fire number prediction model includes:

[0104] Assume that the loss function of the urban fire number forecast model is the mean square error (MSE) of the district-level disaster loss, and the number of training samples is N, where The true and predicted values of the grid map of the cumulative number of fires in the next 30 days for each sample are and , then the loss function is expressed as follows:

[0105] (6)

[0106] The Pytorch framework was used to train the urban fire number prediction model. The Adam optimizer was used as the optimization algorithm, the learning rate was 0.0005, and the number of simulation training iterations was 1000. After each round of training, the validation set was used to calculate the loss function of the urban fire number prediction model. When the number of training iterations of the urban fire number prediction model was reached, the training was stopped; otherwise, the urban fire number prediction model was continued to be trained. During the iteration cycle, the urban fire number prediction model with the best prediction result on the validation set was selected as the final trained urban fire number prediction model, and the parameters of the urban fire number prediction model were determined.

[0107] Finally, the test set is used to evaluate the performance of the urban fire number prediction model. Take 6 samples in the test set as an example. Figure 4 As shown in the figure, the comparison results between the predicted value and the actual value obtained by using the city fire number forecast model are shown. Figure 4 In the data, darker colors indicate a higher number of fires expected in the next 30 days, while lighter colors indicate a lower number of fires. The results demonstrate that the method of the present invention can effectively capture high-risk areas for fires in the next 30 days, demonstrating that the urban fire forecast model of the present invention has high prediction accuracy and is well-suited for implementation.

[0108] The above only describes the basic principles and preferred embodiments of the present invention. Those skilled in the art may make many changes and improvements based on the above description, and these changes and improvements should fall within the scope of protection of the present invention.

Claims

1. A method for predicting the spatial distribution of urban fires based on a deep learning algorithm, characterized in that: The steps include: S1. Collecting original data: Collect historical fire event data in the urban study area, as well as urban land use type data and meteorological factor data corresponding to each historical fire event. The historical fire event data in the urban study area includes the location and time of occurrence of each fire; S2. Preprocessing the raw data collected in step S1: Dividing the urban study area into grid cells, and performing spatiotemporal aggregation processing on the historical fire event data of the urban study area, specifically including: in terms of time, performing daily statistics to divide the data into a daily fire count dataset; in terms of space, dividing the grid cells according to a preset spatial resolution, and counting the geographical coordinate points of the daily fire count into the grid cells to form a daily fire count spatial grid map; for urban land use types, calculating the area ratio of each land use type in each grid; for meteorological factors, averaging the meteorological factor data recorded by all meteorological stations in the urban study area; S3, based on the data pre-processed in step S2, determine the input variables and output variables for the urban fire number forecast model: the input variables include urban land type raster data, past Time series data of meteorological factors of the day and past The three types of input variables are spatial type data, time type data and spatiotemporal type data; the output variable is the future Daily cumulative fire number raster map; , ,and and All are integers; S4. Constructing a city fire number prediction model based on a deep neural network: The city fire number prediction model includes an encoder and a decoder, wherein the encoder is used to process time series variables, and the decoder is used to process urban land type spatial variables and feature variables; The encoder architecture consists of a long short-term memory network connected to a self-attention mechanism network, and the decoder architecture consists of a fully connected network and several layers of convolutional neural networks; The time series components that the encoder processes include meteorological factor time variables and the time component of the historical daily fire count Specific operations include: The meteorological factor time variable and the time component of the historical daily fire count raster map collection Splicing in the time dimension to obtain a composite time series , composite time series The dimension is , composite time series As the input variable of the encoder, the encoder processes the composite time series Encode, and then use the last moment variable of the self-attention mechanism network output as the feature variable , where r is the number of hidden units in the self-attention mechanism network; Decoder used to process urban land use type spatial variables and characteristic variables Specific operations include: First, the feature variables For preprocessing, a fully connected network is used to Transformed into Secondly, use matrix transformation to Further transformed into variables ; Then, the variable and variables Splicing on the channel dimension to obtain composite spatial variables , composite spatial variable The dimension is ; Input the composite spatial variable into the multi-layer convolutional neural network, and the output of the last layer of convolutional neural network is the future The predicted value of the daily cumulative fire number grid map is set as ; S5. Training and evaluation of the urban fire number forecast model: Arrange the input variables and output variables of step S3 to obtain paired data sets, and divide the data sets into a training set, a validation set, and a test set in chronological order and at a preset ratio. The training set and validation set are used to determine the parameters of the urban fire number forecast model, and the test set is used to evaluate the performance of the urban fire number forecast model.

2. The method for predicting the spatial distribution of urban fires based on a deep learning algorithm according to claim 1 is characterized in that: In step S1, the urban land type data includes two categories: urban and rural land and urban construction land. Among them, urban and rural land is subdivided into construction land and non-construction land, and urban construction land is subdivided into residential land, public management and public service land, commercial service facility land, industrial land, logistics and warehousing land, road and transportation facility land, public utility land, green space and square land according to its function; the meteorological factor data includes the daily average temperature, daily average relative humidity and daily cumulative rainfall recorded by all meteorological observation stations in the urban study area.

3. The method for predicting the spatial distribution of urban fires based on a deep learning algorithm according to claim 1 is characterized in that: In step S2, for meteorological factors, the meteorological factor data recorded by all meteorological stations in the urban study area are averaged, specifically including: For different meteorological factors, the data recorded by all meteorological stations in the urban study area were averaged to obtain the time series data of daily average temperature, daily average relative humidity and daily cumulative rainfall in the urban study area.

4. The method for predicting the spatial distribution of urban fires based on a deep learning algorithm according to claim 2 is characterized in that: In step S3, the urban land type raster data is as follows: The urban land type raster maps are stacked into multi-channel image data in the spatial dimension. Assume that the set of all urban land type raster layers is represented as follows: （1） In formula (1), Represents one of the urban land use type layers, , let the length and width of each raster layer be and , then the variable Dimension size ,Right now, As the input variable of the urban fire number prediction model.

5. The method for predicting the spatial distribution of urban fires based on a deep learning algorithm according to claim 4 is characterized in that: In step S3, the past The time series data of meteorological factors for the day are as follows: The meteorological factors are stacked into multi-feature time series data in the time dimension. Let the time of the day be ,past The time series data of meteorological factors for a day are expressed as follows: （2） In formula (2), Indicates the Time series of meteorological factors of the day, , the meteorological factor time series specifically includes three variables: daily average temperature, daily average relative humidity and daily cumulative rainfall, so the variable The dimension size is ,Right now, As the input variable of the urban fire number prediction model.

6. The method for predicting the spatial distribution of urban fires based on a deep learning algorithm according to claim 5 is characterized in that: In step S3, the past The daily fire count raster map set is as follows: past The daily fire count raster map set is represented as follows: （3） In formula (3), Indicates the Raster map of the number of fires per day, ; The empirical orthogonal function decomposition method is used to transform the spatiotemporal data into Perform principal component decomposition and retain the previous principal components, so that the The cumulative variance explanation rate of the principal components is not less than the preset percentage, as follows: （4） In formula (4), Indicates the number of grids in the daily fire count grid map, Represents the spatial component of the decomposed daily fire count raster map set, Represents the time component of the decomposed daily fire count raster map set, extract As the input variable of the urban fire number prediction model.

7. The method for predicting the spatial distribution of urban fires based on a deep learning algorithm according to claim 6 is characterized in that: In step S3, the future The daily cumulative fire count raster map is as follows: By moving the present day into the future Total days The daily fire count grid map collection is aggregated in time to form the future The daily cumulative number of fires grid map is: （5） In formula (5), Indicates the The daily fire count raster map will be used to The daily cumulative fire number raster map is used as the output variable of the urban fire number prediction model.

8. The method for predicting the spatial distribution of urban fires based on a deep learning algorithm according to claim 1 is characterized in that: In step S5, the training of the urban fire number prediction model specifically includes: Assume that the loss function of the urban fire number forecast model is the mean square error (MSE) of the district-level disaster loss, and the number of training samples is N, where A sample of the future The actual and predicted values of the daily cumulative fire number grid map are and , then the loss function is expressed as follows: （6） Set a number of training iterations, use the training set to backpropagate the mean square error (MSE) of the urban fire occurrence forecast model using the optimizer, and use the validation set to calculate the loss function of the urban fire occurrence forecast model after each round of training. When the number of training iterations of the urban fire occurrence forecast model is reached, stop training; otherwise, continue training the urban fire occurrence forecast model. Within the iteration cycle, select the urban fire occurrence forecast model with the best prediction result on the validation set as the final trained urban fire occurrence forecast model, and determine the parameters of the urban fire occurrence forecast model.

Images