Tornado space-time intelligent prediction method fusing multi-source data and physical information

Abstract

The invention discloses a multi-source data and physical information fused tornado space-time intelligent prediction method, which comprises the steps of collecting observation data of a Doppler weather radar and reanalysis data of a mid-term weather forecast center, performing projection conversion, space alignment, time interpolation and the like of space-time dimensions, and fusing key physical characteristics of the tornado; establishing a space-time convolution model network suitable for tornado multi-scale feature extraction, inputting fusion features and corresponding coordinate information, and outputting a prediction control quantity of tornado space-time evolution; a physical control equation of the atmospheric motion and thermodynamic process is converted into a penalty term, a space-time convolution model loss function is embedded, and physical information constraint is carried out on network parameters; and obtaining tornado prediction control quantities of different time windows in the future, and generating a probability graph of the potential development area of the tornado. Through fusion of multi-source data and a physical mechanism, future tornado event probability is quantified, and reliable data support and scientific basis are provided for meteorological prediction and emergency response.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of extreme weather forecasting, in particular to a tornado spatio-temporal intelligent prediction method fusing multi-source data and physical information. BACKGROUND

[0002] Traditional tornado prediction methods mainly rely on Doppler weather radar, and artificial interpretation is performed by analyzing typical features such as mesocyclones, hook echoes, and tornado vortex signatures (TVS) in radar echo maps. However, such methods have many shortcomings: the recognition result is highly dependent on the professional knowledge and practical experience of the forecaster, and it is difficult to achieve standardized and stable output; the artificial analysis process is time-consuming, and it is difficult to meet the minute-level warning needs of traffic emergency response; due to the complex and variable characteristics of tornadoes, especially in the initial stage of their life cycle, the characteristic signals are weak, and it is difficult to effectively deal with small-scale and short-life tornado events. In addition, traditional tornado recognition methods often have a high false positive rate due to excessive reliance on inherent physical assumptions.

[0003] In recent years, artificial intelligence technologies represented by deep learning have made breakthrough progress in the field of meteorology, providing a new solution for the automated analysis of radar data. Some researchers have attempted to apply convolutional neural networks (CNN) and other models to tornado recognition and have achieved some results. However, tornadoes are extreme weather events, and the lack of data and the extreme imbalance between positive and negative samples make model training difficult, leading to false negatives. In addition, models that rely solely on data fitting may produce prediction results that violate atmospheric physical laws, and their generalization ability is questionable when faced with extreme weather events that have not been seen before. Furthermore, the lack of physical interpretability within the model makes it difficult to gain the full trust of forecasters. Therefore, effectively integrating observation and prediction data from different sources and scales can provide comprehensive input for the model. In addition, using data and physical dual driving methods, the model output not only approximates the observed facts but also follows the basic laws of atmospheric dynamics and thermodynamics. This not only significantly improves the prediction robustness and generalization ability of the model in data-sparse or extreme situations, but also provides a solid foundation for the physical interpretability of the prediction results. SUMMARY

[0004] Technical problem: The present application aims to address the shortcomings of traditional tornado prediction methods in terms of prediction length, prediction accuracy, and intelligent warning. A tornado spatio-temporal intelligent prediction method fusing multi-source data and physical information is proposed. This method fuses Doppler weather radar and ECMWF reanalysis data, establishes a spatio-temporal convolutional network model to extract tornado spatio-temporal evolution features, and optimizes network parameters through embedded physical information. At the same time, a tornado potential generation area probability map is generated to effectively quantify tornado prediction.

[0005] Technical scheme: A tornado spatio-temporal intelligent prediction method fusing multi-source data and physical information comprises the following steps: Step 1, multi-source data collection: the occurrence time, place and intensity of different tornado events are counted, and corresponding Doppler weather radar observation data and ECMWF reanalysis data of the medium-term weather forecasting center are collected; Step 2, feature extraction and fusion: through time and space alignment technology, the weather radar data and ECMWF reanalysis data of different resolutions are uniformly interpolated and projected into high-resolution Cartesian grids, the key physical features of the two are extracted, and are fused into four-dimensional data samples of tornado time steps, height, width and channel number; Step 3, tornado prediction based on spatio-temporal convolution network: the data samples are input into the spatio-temporal convolution network composed of an encoder, a jump connection and a decoder, and finally the tornado prediction control quantity is output; The encoder has a nested architecture of a spatio-temporal convolution layer composed of a convolutional long short-term memory network unit, an activation function and a maximum pooling downsampling, and a coordinate convolution layer and a large kernel selection convolution module, for capturing the spatial structure and time dependence of the tornado; The jump connection adopts an attention gate mechanism to effectively integrate the shallow features of the corresponding layers in the encoder to the decoder; The decoder is based on bilinear upsampling considering coordinate information and convolutional long short-term memory network unit, and gradually restores the spatial resolution of the feature map; Step 4, physical information guided prediction and optimization: based on the obtained prediction control quantity, the physical equation residual is calculated through automatic differentiation technology, and a joint loss function including the control equation loss function, the observation data loss function and the boundary condition loss function is defined, and the spatio-temporal convolution network is optimized based on the joint loss function to predict the future tornado; Step 5, output probability quantization: different time windows are set for the spatio-temporal convolution network, the tornado prediction control quantity at different times in the future is obtained, a series of high-resolution tornado potential development area probability maps are generated based on probability prediction, and the confidence interval is calculated by setting a probability threshold, thereby providing clear early warning decision support for meteorological forecast personnel.

[0006] In step 1, the multi-source data collection step specifically comprises: The historical tornado event records of many years are counted from the national meteorological disaster database, and the important information of the starting and ending time, longitude and latitude coordinates, enhanced Fujita scale, path length and maximum width, and disaster situation of each tornado event is stored in a structured manner; According to the central longitude and latitude of each tornado event , the nearest next-generation weather radar site is queried and determined ; the distance between the two is calculated , ensure that it is within the effective detection radius; distance The calculation formula is as follows: (twenty three) Where, is the radius of the Earth, is the latitude difference, is the longitude difference, and are the sine and cosine functions, respectively. is the inverse sine function; For each tornado event, a data collection time window of one hour before and after the event is defined; within this time and space range, download All available original radar-based data files in the data archive, including physical quantities such as reflectivity factor DBZ, radial velocity VEL, spectral width WIDTH, and dual-polarization products of differential phase shift KDP, correlation coefficient RHOHV, and differential reflectivity ZDR; A rectangular area with the central longitude and latitude of each tornado event as the geometric center was defined. The same time window as the radar data was used to download ECMWF reanalysis data within that time period. The physical quantities included: air temperature at a height of 2 meters at the ground or near-ground layer, dew point temperature at a height of 2 meters, east-west and north-south wind components at a height of 10 meters, sea level pressure, and geopotential height, temperature, east-west / north-south wind components, and relative humidity / specific humidity at the standard isobaric surface layers of 250, 300, 500, 700, and 850 hPa.

[0007] In step 2, feature extraction and fusion specifically include: A Lambert conformal conic projection was used, with the Cartesian grid center set to the latitude and longitude coordinates of the start of each tornado event. The grid coverage and grid resolution were set to form a two-dimensional Cartesian grid. In addition, the time step was set to the frequency of the original radar-based data, and the number of time steps for each sample was defined to include continuous observations for one hour before and after the tornado occurred. Each raw radar data file is quality controlled, including filtering out ground clutter and non-meteorological echoes, and deblurring radial velocity. The polar coordinate radar data that has passed the quality control is interpolated to the Cartesian grid defined above using the Barnes objective analysis method, and the grid value is estimated by performing distance-weighted averaging of adjacent observation points. The core weight function is and grid interpolation G The calculation of is as follows: (twenty four) (25) Where, N is the number of observation points, is the natural exponential function; and Respectively The value and weight of each radar observation point, For each point to The distance between observation points, and is an adjustable parameter that affects the degree of smoothing and response scale, For observation points From 1 to N The traversal summation of ; The low, medium and high layers were selected to generate constant height position plane maps respectively. At each height, the reflectivity factor DBZ, radial velocity VEL, spectrum width WIDTH, ratio differential phase shift KDP, correlation coefficient RHOHV and differential reflectivity ZDR radar physical quantities were extracted as independent feature channels. Among ECMWF reanalysis data of different resolutions, the low-resolution ECMWF reanalysis data are upsampled to a high-resolution Cartesian grid consistent with the radar data using a bilinear interpolation method. The hourly resolution reanalysis data are then subjected to linear time interpolation to generate environmental field base data that are consistent with the radar data. On the interpolated high-resolution grid, key environmental parameters such as convective effective potential energy (CAPE), deep wind vertical shear (BWS), and storm relative helicity (SRH) are calculated as additional characteristic channels. The relevant calculation formulas are as follows: (26) (27) (28) Where, is the acceleration due to gravity, z To specify the height, z LFC and z EL are the free convection altitude and equilibrium altitude, respectively. To test the virtual temperature of the air, It is the virtual temperature of the ambient air; z 1 and z 2 are the heights of the lower and upper layers respectively; and At height z 1 and z The east-west wind speed component at location 2, and At height z 1 and z The north-south wind speed component at location 2; For height zhorizontal wind vector, is an estimated storm movement vector, is a horizontal vorticity vector; For each time step, the feature maps generated by all the multiple height layers from the radar and the environmental parameters of ECMWF are stacked along the channel dimension; the three-dimensional feature maps of height, width and channel of the multiple time steps, stacked channel dimension are stacked again in time sequence, and finally a four-dimensional 4D tensor sample X with a dimension of [time step, height, width, channel number] is formed, and the structure can be represented as: (29) In the formula, is a four-dimensional array, T is a continuous time step, H is a grid height, W is a grid width, C is the total number of feature channels.

[0008] In step 3, the tornado prediction based on the space-time convolution network specifically includes: A tornado space-time convolution network composed of an encoder, a decoder and a jump connection is established; a plurality of 4D data channels composed of fused feature data samples after normalization operation are input, including different radar variables, ECMWF variables, coordinate information and binary mask; Low-level tornado space-time features are extracted by the encoder and gradually down-sampled; different layers are provided with different numbers of convolution kernels for hierarchical feature extraction and dimension transformation; the space-time convolution layer is composed of two consecutive convolution long short-term memory network units, an activation function and a maximum pooling down-sampling; wherein the convolution long short-term memory network unit can extract time series and radar spatial features at the same time, and the calculation formula is as follows: (30) (31) (32) (33) (34) In the formula, the subscript t is a time step, X t is the input tensor at the current time, H t-1 is the hidden state of the previous time step, H t is the hidden state at the current time, C t-1 is the cell state of the previous time step, Ct is the current unit state, They are input gate, forget gate, and output gate respectively. is the convolution kernel weight corresponding to the input gate operation, is the convolution kernel weight corresponding to the forget gate operation, is the convolution kernel weight corresponding to the unit state operation, is the convolution kernel weight corresponding to the output gate operation, 、 is the bias term, is the convolution operation, is element-wise multiplication, σ and tanh are activation functions; Then, the coordinate convolution layer with coordinate information is embedded into the input tensor of the spatiotemporal convolution layer to provide an additional coordinate system space layout; the convolution kernel size in the spatiotemporal convolution module is dynamically adjusted through the large kernel selection convolution module to process the context information differences between different tornado targets. The calculation formula is as follows: (35) In the formula, Y is the output variable, X is the input variable, is the convolution operation, is the activation function, is average pooling, is the maximum pooling operation, is a 1×1×1 convolution, Convert the 2-channel pooling feature into N A spatial attention feature map, is the first Layer features, N is the number of convolution kernels, is the feature map From 1 to N The traversal summation of ; The attention gate AG mechanism is introduced to enable the skip connection to focus more on important tornado features and suppress irrelevant background areas, as follows: (36) (37) Where, is the activation function; are the features of the decoder and encoder respectively, For passing Linear change operation after activation, superscript is a linear mapping dimensional space, the superscript T is the transpose operation; 、 and They are , and linearly changing transpose after activation, , respectively and bias terms; is the attention score, is the feature parameter, is the attention coefficient; The spatial dimension of the tornado spatiotemporal feature map is doubled by establishing a decoder considering the bilinear up-sampling and two spatiotemporal convolution layers, and combining the feature information in the skip connection. Based on the linear activation 3D convolution layer of the output layer, the tornado feature prediction control quantity is obtained; finally, the result value is transmitted to the loss function together with the corresponding real tornado feature quantity.

[0009] The physical information guided prediction and optimization in step 4 specifically includes: According to the output of the spatiotemporal convolution network, the tornado prediction control quantity is generated, including the wind speed field u, v , w , the pressure field p and the temperature field T c , and is expressed in the form of a four-dimensional tensor of time, height, width and channel; The Navier-Stokes equation, continuity equation and thermodynamic energy equation are used as physical constraints to ensure that the tornado prediction result meets the laws of fluid mechanics and thermodynamics; the equation forms of the three are as follows: (38) (39) (40) In the formula, is the partial derivative, t is the time, is the wind speed vector, T c is the temperature, is the pressure, is the density, is the kinematic viscosity coefficient, is the thermal diffusion coefficient, is the external force term, is the heat source term, is the gradient operator; The partial derivatives of the prediction control quantity with respect to time and space are calculated, substituted into the physical constraint equation, and the residual value of each point is obtained and normalized to calculate the root mean square error : (41) where, is the residual function of physical equation, is the prediction point from 1 to N the traversal summation; The root mean square error of the prediction control quantity is compared with the weather radar and ECMWF data : (42) where, is the prediction value, is the observation value, is the observation point from 1 to N the traversal summation; The bias of the difference between the prediction value of the boundary point and the boundary condition is calculated by applying the Dirichlet boundary condition to the prediction area : (43) where, is the boundary prediction value, is the boundary value, is the boundary point from 1 to N the traversal summation; The joint total loss function is defined as: (44) where, are weight coefficients, respectively; the back propagation algorithm is used to optimize the deep neural network parameters based on the total loss function , and the model training is completed.

[0010] In step 5, the output probability quantization specifically includes: According to the meteorological forecast requirements, set the prediction time window and time resolution; add a probability prediction module after the network output layer, use the activation function to scale the prediction control quantity at each time point to the [0, 1] interval, generate the tornado occurrence probability of each grid point, and form a four-dimensional probability tensor of time, height, width, and probability value; Based on historical data, calculate the receiver operating characteristic curve and performance diagram of the probability distribution, optimize the threshold to maximize the critical success index; Generate a binary mask by applying a threshold, marking the potential tornado generation area; use the connected component analysis algorithm to cluster the high probability area to generate the boundary of the continuous potential generation area; apply Gaussian smoothing filtering to retain the main high probability area morphology; The mean value and standard deviation of each grid point of the probability sample are calculated, a confidence interval is constructed, and an error bar is superimposed on the probability map; based on the probability threshold and the confidence, a low, medium and high risk warning level is divided; Finally, a tornado time series prediction probability map is generated, a two-dimensional slice is output according to the ground height, the high probability area boundary, the confidence interval and the warning level are marked, and intuitive warning information is provided for tornado forecast personnel.

[0011] Advantages: After adopting the above scheme, the advantages of the present application are as follows: (1) The present application combines Doppler weather radar and ECMWF reanalysis data, improves the physical correlation and feature rationality of the input data, and effectively avoids the problems of tornado data scarcity and unclear features.

[0012] (2) The present application establishes a spatio-temporal convolution network suitable for tornado prediction, improves the ability of the model to extract spatio-temporal evolution features of tornadoes, and has higher prediction accuracy and good engineering application prospect.

[0013] (3) The present application aggregates physical information and data features, reversely constrains the parameter weight of the convolutional neural network, and still maintains high accuracy and stability under complex environment and data distribution changes.

[0014] (4) The present application uses a probability prediction method, can output the probability distribution of the tornado prediction control quantity, realizes the quantification of the credibility of the prediction result, and provides support for risk assessment and intelligent decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a flowchart of the present application; Figure 2 is a schematic diagram of the fusion of weather radar and ECMWF data; Figure 3 is a schematic diagram of the spatio-temporal convolution network; Figure 4 is a schematic diagram of the probability distribution quantization of the tornado prediction control quantity. DETAILED DESCRIPTION

[0016] The technical solutions and advantages of the present application will be described in detail below with reference to the accompanying drawings.

[0017] As shown in Figure 1 , a tornado spatio-temporal intelligent prediction method fusing multi-source data and physical information includes the following steps: Step 1, multi-source data collection: statistics of the occurrence time, location and intensity of different tornado events, collection of corresponding Doppler weather radar observation data and ECMWF reanalysis data of the Medium Range Weather Forecast Center.

[0018] Step 2, feature extraction and fusion: Through time and space alignment technology, weather radar data and ECMWF reanalysis data of different resolutions are uniformly interpolated and projected into high-resolution Cartesian grids, key physical features are extracted, and four-dimensional data samples of tornado time steps, height, width, and channel number are fused.

[0019] Step 3, tornado prediction based on spatio-temporal convolutional network: input the data sample into the spatio-temporal convolutional network composed of encoder, jump connection and decoder, and finally output the tornado prediction control quantity; The encoder architecture nests the spatio-temporal convolutional layer composed of convolutional long short-term memory network unit, activation function and maximum pooling downsampling, and the coordinate convolutional layer and large kernel selection convolutional module, which are used to capture the spatial structure and time dependence of tornadoes; The jump connection adopts attention gate mechanism to effectively integrate the shallow features of the corresponding layers in the encoder to the decoder; The decoder is based on bilinear upsampling and convolutional long short-term memory network unit considering coordinate information, and gradually restores the spatial resolution of the feature map.

[0020] Step 4, physical information guided prediction and optimization: based on the obtained prediction control quantity, the residual of the physical equation is calculated through automatic differentiation technology, and the joint loss function including the control equation loss function, observation data loss function and boundary condition loss function is defined. Based on the joint loss function, the spatio-temporal convolutional network is optimized reversely to predict the future tornado.

[0021] Step 5, output probability quantization: set different time windows for the spatio-temporal convolutional network, obtain the tornado prediction control quantity at different times in the future, generate a series of high-resolution tornado potential development area probability maps based on probability prediction, and set a probability threshold to calculate its confidence interval, providing clear early warning decision support for meteorological forecast personnel.

[0022] In step 1, the multi-source data collection step specifically includes: From the national meteorological disaster database, the 10-year history of tornado event records is counted, and the starting and ending time (UTC time system), longitude and latitude coordinates, enhanced Fujita scale (EF-Scale), path length and maximum width, disaster situation and other important information of each tornado event are stored in a structured manner; according to the central longitude and latitude of each tornado event , the nearest next-generation weather radar site is queried and determined ; the distance between the two is calculated d , to ensure that the distance is within the effective detection radius; the effective detection radius is 230 km, and the distance d The calculation formula is as follows: ,

[0023] where, r R is the earth radius, taken as 6371 km; is the latitude difference, is the longitude difference, and are the sine and cosine functions, respectively, is the inverse sine function; is defined as a one-hour data collection window before and after each tornado event; within this spatio-temporal range, all available raw radar-based data files are downloaded from the data archive, including the following physical quantities: reflectivity factor DBZ, radial velocity VEL, spectral width WIDTH, and dual-polarization products such as specific differential phase KDP, correlation coefficient RHOHV, differential reflectivity ZDR, etc. is defined as a 3x3 latitude-longitude rectangular region with the center latitude and longitude of each tornado event as the geometric center; using the same time window as the radar data, ECMWF reanalysis data within this time period are downloaded, including the following physical quantities: 2-meter height air temperature, 2-meter height dew point temperature, 10-meter height eastward and northward wind components, sea level pressure, and potential height, temperature, eastward / northward wind components, and relative humidity / specific humidity at the 250, 300, 500, 700, and 850 hPa standard isobaric levels.

[0024] The feature extraction and fusion step in step 2 specifically includes: As shown in Figure 2 Lambert conformal conic projection is used to set the Cartesian grid center point as the starting latitude and longitude coordinates of each tornado event; the grid coverage range is set to 256 km x 256 km, and the grid resolution is set to 1 km x 1 km, finally forming a 256 x 256 two-dimensional Cartesian grid; in addition, the time step is set to 5 minutes to match the frequency of the original radar-based data, and the number of time steps for each sample is defined as 24, i.e., including one hour (24 x 5 minutes) of continuous observations before and after the tornado occurrence; For each raw radar-based data file, quality control is performed, including: filtering of ground clutter and non-meteorological echoes, and de-aliasing of radial velocity; the quality-controlled polar coordinate radar data are interpolated to the aforementioned defined Cartesian grid using the Barnes objective analysis method, and the adjacent observation points are distance-weighted averaged to estimate the grid point value, with the core weight function and the calculation of grid point interpolation G as follows: , where, N is the number of observation points, is the natural exponential function; and​ are the values and weights of the th radar observation point, are the distances from each point to the th observation point, r and are adjustable parameters that affect the smoothing degree and response scale, and are set to 0.25 and 2, respectively, are observation points from 1 to N are summed up by traversing; Select 1 km, 3 km, 5 km as low, medium, high height, respectively, to generate constant height position plan at height above ground; On each height layer, extract DBZ, VEL, WIDTH, KDP, RHOHV and ZDR physical quantities as independent feature channels; For ECMWF reanalysis data with low spatial resolution, bilinear interpolation method is used to upsample to high-resolution Cartesian grid with 256x256 km consistent with radar data; Linear time interpolation is used to generate environmental field base data consistent with radar data 5-minute step; On the interpolated high-resolution grid, calculate key environmental parameters such as convective available potential energy CAPE, deep wind vertical shear BWS and storm relative helicity SRH as additional feature channels, and the related calculation formulas are as follows: , where, is the acceleration of gravity, z is the specified height, z LFC and z EL are the free convection height and equilibrium height, respectively, is the virtual temperature of the test air, is the virtual temperature of the environment air; z 1 and z 2 are the lower and upper layer heights, respectively, 1 km and 3 km; and are the east-west wind speed components at heights z 1 and z 2, and are the north-south wind speed components at heights z 1 and z 2; is the horizontal wind vector at height z , is the estimated storm movement vector, is the horizontal vorticity vector; At each time step, the feature maps generated by all the multiple height layers from the radar and the environmental parameters of ECMWF are stacked along the channel dimension; the 24 time steps, stacked channel dimension, height, width, and channel three-dimensional feature maps are stacked in time order again, and finally a four-dimensional 4D tensor sample X with a dimension of [24, 256, 256, C ] is formed, and the structure can be represented as: , In the formula, is a four-dimensional array, T is a continuous time step, H is a grid height, W is a grid width, C is the total number of feature channels.

[0025] In step 3, the tornado prediction step based on the spatio-temporal convolutional network specifically includes: As shown in Figure 3 , a tornado spatio-temporal convolutional network composed of an encoder, a decoder, and a skip connection is established; a plurality of 4D data channels composed of fused feature data of different times and different heights after normalization operation are input, including different radar variables, ECMWF variables, coordinate information, and binary masks; Low-level tornado spatio-temporal features are extracted by the encoder and gradually down-sampled; the number of convolution kernels is set to 64, 128, 256, 512, and 1024, respectively, for hierarchical feature extraction and dimension transformation; in addition, the spatio-temporal convolutional layer is composed of two consecutive convolutional long short-term memory network units, an activation function, and a maximum pooling down-sampling; among them, the convolutional long short-term memory network can extract time series and radar spatial features at the same time, and its calculation formula is as follows: ,

[0026] In the formula, the subscript t is a time step, X t is the input tensor at the current time, H t-1 is the hidden state at the previous time step, H t is the hidden state at the current time, C t-1 is the cell state at the previous time step, C t is the cell state at the current time, are the input gate, the forget gate, and the output gate, is the convolution kernel weight corresponding to the operation of the input gate, is the convolution kernel weight corresponding to the operation of the forget gate, Convolution kernel weights corresponding to the operation of the unit state, Convolution kernel weights corresponding to the operation of the output gate, 、 Bias term, Convolution operation, Element-wise multiplication, σ And tanh is an activation function; Then, the coordinate convolution layer with coordinate information is embedded into the input tensor of the space-time convolution layer to provide additional coordinate system space layout; the convolution kernel size in the space-time convolution module is dynamically adjusted through the large kernel selective convolution module to effectively handle the context information difference between different tornado targets, and the calculation formula is as follows: , In the formula, Y is an output variable, X is an input variable, Convolution operation, Activation function, Average pooling, Maximum pooling operation, 1×1×1 convolution, Convert 2 channel pooling features into N Spatial attention feature map, The first Layer feature of X, N Convolution kernel number, Feature map From 1 to N The traversal sum; The attention gate AG mechanism is introduced to make the skip connection pay more attention to important tornado features and suppress irrelevant background areas, as follows: , In the formula, Activation function; Respectively, the features of the decoder and the encoder, After linear change operation after activation, Superscript Linear mapping dimension space, superscript T Transpose operation; And Respectively, And Linear change transpose after activation, Respectively, And Bias term; Attention score, Feature parameter, Attention coefficient; The spatial dimension of the tornado spatiotemporal feature map is doubled by establishing a decoder considering the bilinear up-sampling of coordinate information and two standard spatiotemporal convolution layers, and combining the feature information in the skip connection.

[0027] Based on the linear activation 3D convolution layer of the output layer, the tornado feature prediction control quantity is obtained; finally, the result value is transmitted to the loss function together with the corresponding real tornado feature quantity.

[0028] The physical information guided prediction and optimization step specifically includes: According to the spatiotemporal convolution network output, the tornado prediction control quantity is generated, including the wind speed field , the pressure field p and the temperature field T c , and is expressed as a four-dimensional tensor form of time, height, width and channel; The Navier-Stokes equation, continuity equation and thermodynamic energy equation are used as physical constraints to ensure that the tornado prediction result meets the laws of fluid mechanics and thermodynamics; the equation forms of the three are as follows: , In the formula, is the partial derivative, t is the time, is the wind speed vector, T c is the temperature, is the air pressure, is the density, is the kinematic viscosity coefficient, is the thermal diffusion coefficient, is the external force term, is the heat source term, is the gradient operator; The partial derivative of the prediction control quantity with respect to time and space is calculated, substituted into the physical constraint equation, the residual value of each point is obtained, and normalized processing is performed, and the root mean square error : , In the formula, F is the residual function of the physical equation, is the prediction point from 1 to N summed up; The prediction control quantity is compared with the Doppler weather radar and ECMWF observation data, and the root mean square error : , In the formula, is the predicted value, for observation values, for observation points from 1 to N the traversal sum of; a Dirichlet boundary condition is applied to the prediction area, and the deviation between the boundary point prediction value and the boundary condition is calculated : , wherein, is a boundary prediction value, is a boundary value, is a boundary point from 1 to N the traversal sum of; joint total loss function is defined as: , wherein, are weight coefficients respectively; based on the total loss function , the deep neural network parameters are optimized, and the model training is completed.

[0029] In step 5, the output probability quantization step specifically includes: According to the meteorological forecast demand, the prediction time window is set to be 1 hour in the future, and the time resolution is 5 minutes; as Figure 4 shown, a probability prediction module is added after the network output layer, using an activation function to map the prediction control amount at each time point to the [0, 1] interval, generating the tornado occurrence probability of each grid point, forming a four-dimensional probability tensor of time, height, width and probability value; Based on historical data, the receiver operating characteristic curve and performance graph of the probability distribution are calculated, and the threshold value is optimized to maximize the critical success index; By applying the threshold value to generate a binary mask, the potential tornado generation area is marked; the connected domain analysis algorithm is used to cluster the high probability area to generate the boundary of the continuous potential generation area; Gaussian smoothing filter is applied to retain the main high probability area form; The mean and standard deviation of each grid point of the probability sample are calculated, and a 95% confidence interval is constructed, and the confidence interval is superimposed on the probability graph in the form of error bars; based on the probability threshold and the confidence, the low, medium and high risk warning levels are divided; Finally, the tornado time series prediction probability graph is generated, and the two-dimensional slices are output according to the ground height, the high probability area boundary, the confidence interval and the warning level are marked, and intuitive warning information is provided for tornado forecast personnel.

Claims

1. A method for intelligent spatiotemporal prediction of tornadoes that integrates multi-source data and physical information, characterized by: The following steps are involved: Step 1: Multi-source data collection: Count the occurrence time, location, and intensity of different tornado events, and collect corresponding Doppler weather radar observation data and ECMWF reanalysis data; Step 2, feature extraction and fusion: Using temporal and spatial alignment techniques, weather radar data and ECMWF reanalysis data of different resolutions are uniformly interpolated and projected onto a high-resolution Cartesian grid. Key physical features of both are extracted and fused into a four-dimensional data sample of tornado time steps, height, width, and number of channels. Step 3, tornado prediction based on spatiotemporal convolutional network: input the data sample into the spatiotemporal convolutional network composed of an encoder, a skip connection, and a decoder, and finally output the tornado prediction control quantity; The encoder architecture nests a spatiotemporal convolution layer consisting of a convolutional long short-term memory network unit, an activation function, and maximum pooling downsampling, as well as a coordinate convolution layer and a large kernel selection convolution module to capture the spatial structure and temporal dependency of the tornado; The jump connection adopts the attention gate mechanism to effectively integrate the shallow features of the corresponding level in the encoder into the decoder; The decoder gradually restores the spatial resolution of the feature map based on bilinear upsampling and convolutional long short-term memory network units considering coordinate information; Step 4: Physical information-guided prediction and optimization: Based on the predicted control variables, the residuals of the physical equations are calculated using automatic differentiation techniques. A joint loss function is defined that includes the control equation loss function, the observation data loss function, and the boundary condition loss function. This joint loss function is then used to inversely optimize the spatiotemporal convolutional network to predict future tornadoes. Step 5: Output probability quantification: Set different time windows for the spatiotemporal convolutional network to obtain tornado prediction control quantities at different moments in the future. Based on the probability prediction, a series of high-resolution probability maps of potential tornado development areas are generated. The probability threshold is set to calculate its confidence interval, providing clear warning decision support for meteorological forecasters.

2. The method for intelligent spatiotemporal prediction of tornadoes by integrating multi-source data and physical information according to claim 1, characterized in that: In step 1, the multi-source data collection step specifically includes: The national meteorological disaster database collects historical tornado event records over the years and stores important information such as the start and end time, longitude and latitude coordinates, enhanced Fujita scale, path length and maximum width, and disaster situation of each tornado event in a structured manner. Based on the central latitude and longitude of each tornado event , query and determine the nearest next generation weather radar site ; Calculate the distance between the two , ensure that it is within the effective detection radius; distance The calculation formula is as follows: （1） Where, is the radius of the Earth, is the latitude difference, is the longitude difference, and are the sine and cosine functions, respectively. is the inverse sine function; For each tornado event, a data collection time window of one hour before and after the event is defined; within this time and space range, download All available original radar-based data files in the data archive, including physical quantities such as reflectivity factor DBZ, radial velocity VEL, spectral width WIDTH, relative differential phase shift KDP, correlation coefficient RHOHV, and dual-polarization product of differential reflectivity ZDR; A rectangular area with the central longitude and latitude of each tornado event as the geometric center was defined. The same time window as the radar data was used to download ECMWF reanalysis data within that time period. The physical quantities included: air temperature at a height of 2 meters at the ground or near-ground layer, dew point temperature at a height of 2 meters, east-west and north-south wind components at a height of 10 meters, sea level pressure, and geopotential height, temperature, east-west / north-south wind components, and relative humidity / specific humidity at the standard isobaric surface layers of 250, 300, 500, 700, and 850 hPa.

3. The method for intelligent spatiotemporal prediction of tornadoes by integrating multi-source data and physical information according to claim 2, characterized in that: In step 2, feature extraction and fusion specifically include: The Lambert conformal conic projection is used to set the center point of the Cartesian grid to the starting latitude and longitude coordinates of each tornado event; the grid coverage and grid resolution are set to form a two-dimensional Cartesian grid; in addition, the time step is set to the frequency of the original radar-based data, and the number of time steps for each sample is defined to include continuous observations for one hour before and after the tornado occurs; each original radar-based data file is quality controlled, including: filtering of ground clutter and non-meteorological echoes, and deblurring of radial velocity; the polar coordinate radar data that have passed the quality control are interpolated to the Cartesian grid defined above using the Barnes objective analysis method, and the grid point value is estimated by performing distance-weighted averaging of adjacent observation points, with the core weight function and grid interpolation G The calculation of is as follows: （2） （3） Where, N is the number of observation points, is the natural exponential function; and Respectively The value and weight of each radar observation point, For each point to The distance between observation points, and is an adjustable parameter that affects the degree of smoothing and response scale, For observation points From 1 to N The traversal summation of ; The low, medium and high layers were selected to generate constant height position plane maps respectively. At each height, the reflectivity factor DBZ, radial velocity VEL, spectrum width WIDTH, ratio differential phase shift KDP, correlation coefficient RHOHV and differential reflectivity ZDR radar physical quantities were extracted as independent feature channels. Among ECMWF reanalysis data of different resolutions, the low-resolution ECMWF reanalysis data are upsampled to a high-resolution Cartesian grid consistent with the radar data using a bilinear interpolation method. The hourly resolution reanalysis data are then subjected to linear time interpolation to generate environmental field base data that are consistent with the radar data. On the interpolated high-resolution grid, key environmental parameters such as convective effective potential energy (CAPE), deep wind vertical shear (BWS), and storm relative helicity (SRH) are calculated as additional characteristic channels. The relevant calculation formulas are as follows: （4） （5） （6） Where, is the acceleration due to gravity, z To specify the height, z LFC and z EL are the free convection altitude and equilibrium altitude, respectively. To test the virtual temperature of the air, It is the virtual temperature of the ambient air; z 1 and z 2 are the heights of the lower and upper layers respectively; and At height z 1 and z The east-west wind speed component at location 2, and At height z 1 and z The north-south wind speed component at location 2; For height z The horizontal wind vector at is the estimated storm motion vector, is the horizontal vorticity vector; For each time step, all feature maps generated by multiple altitude layers from the radar and the environmental parameters of ECMWF are stacked along the channel dimension; the height, width, and channel three-dimensional feature maps of multiple time steps and the stacked channel dimensions are stacked again in chronological order, and finally a 4D tensor sample X with dimensions of [time steps, height, width, number of channels] is formed. Its structure can be expressed as: （7） Where, is a four-dimensional array, T is the number of consecutive time steps, H is the grid height, W is the grid width, C is the total number of feature channels.

4. The method for intelligent spatiotemporal prediction of tornadoes by integrating multi-source data and physical information according to claim 3, characterized in that: In step 3, the tornado prediction based on the spatiotemporal convolutional network specifically includes: A tornado spatiotemporal convolutional network consisting of an encoder, a decoder, and skip connections was established. The input was multiple 4D data channels consisting of normalized fused feature data samples, including different radar variables, ECMWF variables, coordinate information, and binary masks. The encoder extracts low-level tornado spatiotemporal features and gradually downsamples them. Different layers are assigned different numbers of convolution kernels for hierarchical feature extraction and dimensionality transformation. The spatiotemporal convolution layer consists of two consecutive convolutional long short-term memory (LSTM) network units, an activation function, and a maximum pooling downsampling operation. The convolutional LSTM network unit can simultaneously extract time series and radar spatial features. The calculation formula is as follows: （8） （9） （10） （11） （12） In the formula, the subscript t is the number of time steps, X t Input tensor for the current moment, H t-1 is the hidden state at the previous time step, H t The hidden state at the current moment, C t-1 is the cell state at the previous time step, C t is the current unit state, They are input gate, forget gate, and output gate respectively. is the convolution kernel weight corresponding to the input gate operation, is the convolution kernel weight corresponding to the forget gate operation, is the convolution kernel weight corresponding to the unit state operation, is the convolution kernel weight corresponding to the output gate operation, 、 is the bias term, is the convolution operation, is element-wise multiplication, σ and tanh are activation functions; Then, the coordinate convolution layer with coordinate information is embedded into the input tensor of the spatiotemporal convolution layer to provide an additional coordinate system space layout; the convolution kernel size in the spatiotemporal convolution module is dynamically adjusted through the large kernel selection convolution module to process the context information differences between different tornado targets. The calculation formula is as follows: （13） In the formula, Y is the output variable, X is the input variable, is the convolution operation, is the activation function, is average pooling, is the maximum pooling operation, is a 1×1×1 convolution, Convert the 2-channel pooling feature into N A spatial attention feature map, is the first Layer features, N is the number of convolution kernels, is the feature map From 1 to N The traversal summation of ; The attention gate AG mechanism is introduced to enable the skip connection to focus more on important tornado features and suppress irrelevant background areas, as follows: （14） （15） Where, is the activation function; are the features of the decoder and encoder respectively, For passing Linear change operation after activation, superscript is a linear mapping dimensional space, the superscript T is the transpose operation; and They are and Linear change transpose after activation; They are and The bias term; Score attention. is the characteristic parameter, is the attention coefficient; By building a decoder with bilinear upsampling and two spatiotemporal convolutional layers that consider coordinate information and combining feature information from skip connections, the spatial dimension of the tornado spatiotemporal feature map is doubled. Based on the linear activation 3D convolution layer of the output layer, the tornado feature prediction control quantity is obtained; finally, the result value together with the corresponding real tornado feature quantity is passed to the loss function.

5. The method for intelligent spatiotemporal prediction of tornadoes by integrating multi-source data and physical information according to claim 4, characterized in that: In step 4, physical information-guided prediction and optimization specifically include: According to the output of the spatiotemporal convolutional network, tornado prediction control quantities are generated, including wind speed field , pressure field p and temperature field T c , and expressed as a four-dimensional tensor of time, height, width, and channel; The Navier-Stokes equations, continuity equation, and thermodynamic energy equation are used as physical constraints to ensure that tornado prediction results meet the laws of fluid mechanics and thermodynamics. The three equations are in the following form: （16） （17） （18） Where, is the partial derivative, t is the number of time steps, is the wind speed vector, T c is the temperature, is the air pressure, is the density, is the kinematic viscosity coefficient, is the thermal diffusivity, is the external force term, is the heat source term, is the gradient operator; Calculate the partial derivatives of the predicted control quantity with respect to time and space, substitute them into the physical constraint equation, obtain the residual value of each point, perform normalization, and calculate its root mean square error : （19） Where, is the residual function of the physical equation, Forecast point From 1 to N The traversal summation of ; Compare the predicted control quantity with weather radar and ECMWF data and calculate their root mean square error : （20） Where, is the predicted value, is the observed value, For observation points From 1 to N The traversal summation of ; Apply Dirichlet boundary conditions to the prediction region and calculate the deviation of the difference between the predicted value at the boundary point and the boundary condition : （21） Where, is the boundary prediction value, is the boundary value, For boundary points From 1 to N The traversal summation of ; Joint total loss function Defined as: （22） Where, are weight coefficients respectively; the back propagation algorithm is used based on the total loss function , optimize the deep neural network parameters and complete the model training.

6. The method for intelligent spatiotemporal prediction of tornadoes by integrating multi-source data and physical information according to claim 5, characterized in that: In step 5, the output probability quantization specifically includes: setting the prediction time window and time resolution according to the weather forecast requirements; adding a probability prediction module after the network output layer, using an activation function to scale the prediction control quantity at each time point to the interval [0, 1], generating the probability of tornado occurrence at each grid point, and forming a four-dimensional probability tensor of time, height, width, and probability value; Based on historical data, the receiver operating characteristic curve and performance graph of the probability distribution are calculated, and the threshold is optimized to maximize the key success index; A binary mask is generated by applying a threshold to mark potential tornado formation areas; a connected domain analysis algorithm is used to cluster high-probability areas and generate continuous potential formation area boundaries; and a Gaussian smoothing filter is applied to preserve the morphology of the main high-probability areas. Calculate the mean and standard deviation of each grid point in the probability sample, construct a confidence interval, and overlay it on the probability map in the form of error bars; based on the probability threshold and confidence level, divide the risk warning level into low, medium, and high levels; Finally, a tornado time series prediction probability map is generated, and two-dimensional slices are output according to the ground height, marking the boundaries of high-probability areas, confidence intervals and warning levels, providing tornado forecasters with intuitive warning information.

Images