Tornado spatio-temporal intelligent prediction method fusing multi-source data and physical information

Abstract

The application discloses a tornado spatio-temporal intelligent prediction method fusing multi-source data and physical information, and comprises the following steps: collecting observation data of Doppler weather radar and reanalysis data of a medium-term weather forecast center, performing projection transformation, spatial alignment and time interpolation operations in the time-space dimension, and fusing key physical characteristics of tornadoes; a spatio-temporal convolution model network suitable for multi-scale feature extraction of tornadoes is established, input fusion characteristics and corresponding coordinate information are input, and a prediction control amount of spatio-temporal evolution of tornadoes is output; a physical control equation of atmospheric movement and thermodynamics is converted into a penalty term, and is embedded into a spatio-temporal convolution model loss function to constrain network parameters with physical information; tornado prediction control amounts of different time windows in the future are obtained, and a tornado potential development area probability graph is generated. Through fusion of multi-source data and physical mechanisms, a 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 solution: A tornado spatiotemporal intelligent prediction method fusing multi-source data and physical information comprises the following steps:

[0006] Step 1, multi-source data collection: the occurrence time, location 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;

[0007] 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 both are extracted, and are fused into four-dimensional data samples of tornado time steps, height, width and channel number;

[0008] Step 3, tornado prediction based on spatiotemporal convolution network: the data samples are input into the spatiotemporal convolution network composed of an encoder, a jump connection and a decoder, and finally the tornado prediction control quantity is output;

[0009] The encoder has a nested architecture of a spatiotemporal 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 tornadoes;

[0010] The jump connection adopts an attention gate mechanism to effectively integrate the shallow features of the corresponding layers in the encoder to the decoder;

[0011] 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;

[0012] 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 a control equation loss function, an observation data loss function and a boundary condition loss function is defined, and the spatiotemporal convolution network is optimized based on the joint loss function to predict future tornadoes;

[0013] Step 5, output probability quantization: different time windows are set for the spatiotemporal 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.

[0014] In step 1, the multi-source data collection step specifically comprises:

[0015] From the national meteorological disaster database, the historical tornado event records are counted and the important information of each tornado event, such as the start and end time, longitude and latitude coordinates, enhanced Fujita scale, path length and maximum width, and disaster situation, are structured and stored;

[0016] 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 , ensuring that it is within the effective detection radius; the distance The calculation formula is as follows:

[0017] (23)

[0018] In the formula, is the radius of the earth, is the latitude difference, is the longitude difference, and are the sine and cosine functions respectively, is the arcsine function;

[0019] A one-hour data collection time window before and after each tornado event is defined; within this spatiotemporal range, all available raw radar-based data files in the Data archive are downloaded, including dual-polarization products such as reflectivity factor DBZ, radial velocity VEL, spectral width WIDTH, specific differential phase KDP, correlation coefficient RHOHV, and differential reflectivity ZDR;

[0020] With the central longitude and latitude of each tornado event as the geometric center, a rectangular region with latitude and longitude as the range is defined; using the same time window as the radar data, ECMWF reanalysis data within this time period are downloaded, including physical quantities such as 2-meter height air temperature, 2-meter height dew point temperature, 10-meter height east-west and north-south wind components, sea level pressure, and standard isobaric surface layer potential height, temperature, east-west / north-south wind components, and relative humidity / specific humidity at 250, 300, 500, 700, and 850 hPa.

[0021] In step 2, feature extraction and fusion specifically includes:

[0022] Lambert conformal conic projection is used to set the Cartesian grid center point as the starting longitude and latitude coordinates of each tornado event; the grid coverage range and grid resolution are set, and finally a two-dimensional Cartesian grid is formed; in addition, the time step is set to the frequency of the original radar-based data, and the time step number of each sample is defined to include one hour of continuous observation before and after the tornado occurs.

[0023] For each original radar-based data file, quality control is performed, including: filtering of ground clutter and non-meteorological echoes, de-aliasing of radial velocity; the quality-controlled polar coordinate radar data is interpolated to the aforementioned defined Cartesian grid by the Barnes objective analysis method, and the grid point values are estimated by distance weighted averaging of adjacent observation points, the core weight function and grid interpolation G is calculated as follows:

[0024] (24)

[0025] (25)

[0026] wherein, N is the number of observation points, is a natural exponential function; and are the value and weight of the i-th radar observation point, is the distance from each point to the i-th observation point, and are adjustable parameters affecting the degree of smoothing and response scale, is the traversal summation of observation points from 1 to . N Three layer heights of low, medium and high are selected, respectively generating constant height position plan; on each height, the reflectivity factor DBZ, radial velocity VEL, spectral width WIDTH, specific differential phase shift KDP, correlation coefficient RHOHV and differential reflectivity ZDR radar physical quantities are extracted as independent feature channels;

[0027] In different resolution ECMWF reanalysis data, the bilinear interpolation method is used for spatial low resolution ECMWF reanalysis data, which is up-sampled to the high resolution Cartesian grid consistent with the radar data; the hourly resolution reanalysis data is generated by linear time interpolation, which is consistent with the radar data step consistent with the environment field base data; on the interpolated high resolution grid, the key environmental parameters of convective available potential energy CAPE, deep wind vertical shear BWS and storm relative helicity SRH are calculated as additional feature channels, and the related calculation formulas are as follows:

[0028]

[0029] (26)

[0030] (27)

[0031] (28)

[0032] wherein,​​ g is the acceleration of gravity, z h is the specified height, z LFC and z EL fc and he are the free convection height and equilibrium height, respectively, T0 is the virtual temperature of the test air, T is the virtual temperature of the ambient air; z 1 and z 2 are the heights of the lower and upper layers, respectively; and are the eastward 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;

[0033] 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:

[0034] (29)

[0035] 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.

[0036] In step 3, the tornado prediction based on the spatio-temporal convolution network specifically includes:

[0037] A tornado spatio-temporal 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 masks;

[0038] The low-level tornado spatio-temporal 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 spatio-temporal convolution layer is composed of two consecutive convolution long short-term memory network units, an activation function and a maximum pooling down-sampling; among them, 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:

[0039] (30)

[0040] (31)

[0041] (32)

[0042] (33)

[0043] (34)

[0044] In the formula, the subscript t is the time step, X t is the input tensor at the current moment, H t-1 is the hidden state at the previous time step, H t is the hidden state at the current moment, C t-1 is the cell state at the previous time step, C t is the cell state at the current moment, are the input gate, the forget gate and the output gate respectively, 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, is the convolution kernel weight corresponding to the operation of the cell state, is the convolution kernel weight corresponding to the operation of the output gate, , is the bias term, is the convolution operation, is the element-wise multiplication, σ and tanh are activation functions;

[0045] Then, the coordinate convolution layer with coordinate information is embedded into the input tensor of the spatio-temporal convolution layer to provide additional spatial layout of the coordinate system; the convolution kernel size in the spatio-temporal convolution module is dynamically adjusted through the large kernel selection convolution module to handle the differences in context information between different tornado targets, and the calculation formula is as follows:

[0046] (35)

[0047] where Y is the output variable, X is the input variable, is a convolution operation, is an activation function, is an average pooling, is a max pooling operation, is a 1x1x1 convolution, converts 2-channel pooling features into N spatial attention feature maps, is the i-th layer feature of X, N is the number of convolution kernels, is the feature map from 1 to N traversal sum;

[0048] The attention gate AG mechanism is introduced to enable the skip connection to pay more attention to important tornado features and suppress irrelevant background areas, as follows:

[0049] (36)

[0050] (37)

[0051] where, is an activation function; are the features of the decoder and the encoder, respectively, is a linear change operation after activation, and the superscript is the linear mapping dimension space, and the superscript is the transpose operation; T , and are the linear change transpose of , and activation, , are the bias terms of and ; 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 with bilinear upsampling considering coordinate information and two spatiotemporal convolution layers, combined with feature information in the skip connection.

[0052] 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.

[0053] ​

[0054] The step 4, the physical information guided prediction and optimization specifically includes:

[0055] According to the spatio-temporal convolution network output, a tornado prediction control quantity is generated, including a wind speed field u, v , w a pressure field p and a temperature field T c , and is expressed in the form of a four-dimensional tensor of time, height, width and channel;

[0056] 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:

[0057] (38)

[0058] (39)

[0059] (40)

[0060] In the formula, is a partial derivative, t is time, is a wind speed vector, T c is temperature, is pressure, is density, is a kinematic viscosity coefficient, is a thermal diffusion coefficient, is an external force term, is a heat source term, is a gradient operator;

[0061] 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 is calculated:

[0062] (41)

[0063] In the formula, is the residual function of the physical equation, is the traversal sum from 1 to of the prediction point N ;

[0064] The prediction control quantity is compared with the weather radar and ECMWF data, and the root mean square error is calculated:

[0065] (42)

[0066] wherein, is the predicted value, is the observed value, is the observation point is the traversal sum from 1 to N ; and

[0067] a Dirichlet boundary condition is applied to the prediction area, and a deviation of a difference between a boundary point prediction value and the boundary condition is calculated :

[0068] (43)

[0069] wherein, is the boundary prediction value, is the boundary value, is the boundary point is the traversal sum from 1 to N ; and

[0070] a joint total loss function is defined as:

[0071] (44)

[0072] wherein, are weight coefficients, respectively; a back propagation algorithm is used to optimize deep neural network parameters based on the total loss function , and model training is completed.

[0073] In the step 5, the output probability quantization specifically includes:

[0074] According to the meteorological forecast requirement, a prediction time window and a time resolution are set; a probability prediction module is added after a network output layer, an activation function is used to scale a prediction control amount of each time point to a [0, 1] interval, a tornado occurrence probability of each grid point is generated, and a four-dimensional probability tensor of time, height, width and probability value is formed;

[0075] Based on historical data, a receiver operating characteristic curve and a performance graph of the probability distribution are calculated, and a threshold value is optimized to maximize a critical success index;

[0076] A binary mask is generated by applying the threshold value, and a potential tornado generation area is marked; a connected domain analysis algorithm is used to cluster a high probability area, and a continuous potential generation area boundary is generated; a Gaussian smoothing filter is applied to retain a main high probability area form;

[0077] 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 graph in the form of an error bar; based on the probability threshold and the confidence, a low, medium and high risk warning level is divided;

[0078] Finally, a tornado time series prediction probability graph 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.

[0079] Advantages: After adopting the above scheme, the advantages of the present application are as follows:

[0080] (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.

[0081] (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.

[0082] (3) The present application aggregates physical information and data features, and reversely constrains the parameter weight of the convolutional neural network, which can maintain high accuracy and stability under complex environment and data distribution changes.

[0083] (4) The present application uses a probability prediction method to output the probability distribution of the tornado prediction control variable, realizes the quantification of the credibility of the prediction result, and provides support for risk assessment and intelligent decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0084] Figure 1 is a flowchart of the present application;

[0085] Figure 2 is a schematic diagram of the fusion of weather radar and ECMWF data;

[0086] Figure 3 is a schematic diagram of the spatio-temporal convolution network;

[0087] Figure 4 is a schematic diagram of the probability distribution quantization of the tornado prediction control variable. DETAILED DESCRIPTION

[0088] The technical solutions and advantages of the present application will be described in detail below with reference to the accompanying drawings.

[0089] As shown in Figure 1 , a tornado spatio-temporal intelligent prediction method fusing multiple source data and physical information includes the following steps:

[0090] Step 1, multi-source data collection: collect the occurrence time, location and intensity of different tornado events, and collect the corresponding Doppler weather radar observation data and ECMWF reanalysis data of the Medium Range Weather Forecast Center.

[0091] 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 grid, the key physical features of the two are extracted, and they are fused into four-dimensional data samples of tornado time steps, height, width and channel number.

[0092] 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;

[0093] 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 tornado;

[0094] The jump connection adopts attention gate mechanism to effectively integrate the shallow features of the corresponding layers in the encoder to the decoder;

[0095] The decoder is based on bilinear upsampling considering coordinate information and convolutional long short-term memory network unit, which gradually restores the spatial resolution of the feature map.

[0096] Step 4, physical information guided prediction and optimization: based on the obtained prediction control quantity, the residual of physical equation is calculated through automatic differentiation technology, and the joint loss function including control equation loss function, observation data loss function and boundary condition loss function is defined, and the spatio-temporal convolutional network is optimized based on the joint loss function to predict the future tornado.

[0097] Step 5, output probability quantization: set different time windows for the spatio-temporal convolutional network, obtain the tornado prediction control quantity at different time, generate a series of high-resolution tornado potential development area probability maps based on probability prediction, and set the probability threshold to calculate the confidence interval, providing clear early warning decision support for meteorological forecast personnel.

[0098] In step 1, the multi-source data collection step specifically includes:

[0099] From the national meteorological disaster database, the 10-year history of tornado event records is counted, and the important information of each tornado event such as the start and end time (UTC system), longitude and latitude coordinates, enhanced Fujita scale (EF-Scale), path length and maximum width, disaster situation, etc. is stored in a structured manner; according to the central longitude and latitude of each tornado event , query and determine the nearest next-generation weather radar site ; calculate the distance between them d , ensuring that the distance is within the effective detection radius; the effective detection radius is taken as 230 km, the distance d The formula is as follows:

[0100] ,

[0101] In the formula, r is the radius of the earth, 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;

[0102] Define a one-hour data acquisition time window before and after each tornado event; within this spatiotemporal range, download all available original radar-based data files in 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, and differential reflectivity ZDR;

[0103] Take the central longitude and latitude of each tornado event as the geometric center to define a 3x3 longitude and latitude rectangular region; using the same time window as the radar data, download the ECMWF reanalysis data within this time period, including the following physical quantities: 2-meter height air temperature, 2-meter height dew point temperature, 10-meter height east-west and north-south wind components, sea level pressure, and standard isobaric surface layer potential height, temperature, east-west / north-south wind components, and relative humidity / specific humidity at 250, 300, 500, 700, and 850 hPa.

[0104] In step 2, the feature extraction and fusion step specifically includes:

[0105] As Figure 2 shown, using Lambert conformal conic projection, set the Cartesian grid center point as the starting longitude and latitude 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 time step number of each sample is defined as 24, i.e., including one hour (24 x 5 minutes) of continuous observation before and after the tornado occurs;

[0106] For each original radar-based data file, quality control is performed, including: filtering of ground clutter and non-meteorological echoes, de-aliasing of radial velocity; the quality-controlled polar coordinate radar data is interpolated to the aforementioned defined Cartesian grid by Barnes objective analysis method, and the grid point values are estimated by distance weighted average of adjacent observation points, the core weight function and grid interpolation G is calculated as follows:

[0107] ,

[0108] where, N is the number of observation points, is a natural exponential function; and are the value and weight of the th radar observation point, is the distance from each point to the th observation point, r and are adjustable parameters that affect the degree of smoothing and response scale, respectively taking 0.25 and 2, is the traversal summation of observation points from 1 to N .

[0109] Select 1 km, 3 km, 5 km as low, medium and high layer 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;

[0110] For ECMWF reanalysis data with low spatial resolution, bilinear interpolation method is used to upsample to high-resolution Cartesian grid with 256×256 km consistent with radar data; linear time interpolation is used to generate environmental field base data consistent with 5-minute step of radar data for hourly resolution reanalysis data; on the interpolated high-resolution grid, key environmental parameters such as convective available potential energy CAPE, deep wind vertical shear BWS and storm relative helicity SRH are calculated as additional feature channels, and the related calculation formulas are as follows:

[0111] ,

[0112] 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, virtual temperature of the ambient air; z 1 and z 2 are the height of the lower and upper layers, respectively, taken as 1 km and 3 km; and are the eastward wind speed components at the heights z 1 and z 2, and are the north-south wind speed components at the heights z 1 and z 2; is the horizontal wind vector at the height z , is the estimated storm movement vector, is the horizontal vorticity vector;

[0113] For each time step, the feature maps generated by all the multiple height layers from the radar and the environmental parameters of the ECMWF are stacked along the channel dimension; the 24 time steps, the stacked height, width, and channel three-dimensional feature maps are stacked again in time order, 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:

[0114] ,

[0115] 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.

[0116] In the step 3, the tornado prediction based on the spatio-temporal convolution network specifically includes:

[0117] As shown in Figure 3 , a tornado spatio-temporal convolution network composed of an encoder, a decoder, and a skip connection is established; a plurality of 4D data channels composed of fusion feature data of different times and different heights after normalization operation are input, including different radar variables, ECMWF variables, coordinate information, and binary masks;

[0118] The 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 convolution layer is composed of two consecutive convolution long short-term memory network units, an activation function and a maximum pooling down-sampling; among them, the convolution long short-term memory network can extract time series and radar spatial features at the same time, and its calculation formula is as follows:

[0119] ,

[0120] In the formula, the subscript t is the 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 respectively, 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, is the convolution kernel weight corresponding to the operation of the cell state, is the convolution kernel weight corresponding to the operation of the output gate, , is the bias term, is the convolution operation, is the element-wise multiplication, σ and tanh are activation functions;

[0121] Then, the coordinate convolution layer with coordinate information is embedded into the input tensor of the spatio-temporal convolution layer to provide additional spatial layout of the coordinate system; the convolution kernel size in the spatio-temporal convolution module is dynamically adjusted through the large kernel selective convolution module, which effectively handles the differences in context information between different tornado targets, and its calculation formula is as follows:

[0122] ,

[0123] In the formula, Y is the output variable, X is the input variable, is the convolution operation, is the activation function, is the average pooling, is the maximum pooling operation, is the 1x1x1 convolution, converts 2-channel pooling features intoN A spatial attention feature map, For X's Layer features, N The number of convolution kernels, For feature map Take from 1 N Iterate through and sum the results;

[0124] An attention-gating (AG) mechanism is introduced to enable skip connections to focus more on important tornado features and suppress irrelevant background regions, as detailed below:

[0125] ,

[0126] In the formula, For activation functions; These are the features of the decoder and encoder, respectively. For the process Linear transformation operation after activation, superscript For a linearly mapped dimensional space, the superscript... T This is a transpose operation; and They are respectively and The linear transformation after activation is transposed. They are respectively and The bias term; Rate attention For feature parameters, Note the coefficient;

[0127] By establishing a decoder that considers coordinate information through bilinear upsampling and two standard spatiotemporal convolutional layers, and combining feature information from skip connections, the spatial dimension of the tornado spatiotemporal feature map is doubled.

[0128] A linearly activated 3D convolutional layer based on the output layer is used to obtain the tornado feature prediction control quantity; finally, the result value along with the corresponding real tornado feature quantity is passed to the loss function.

[0129] Step 4, the physical information-guided prediction and optimization step specifically includes:

[0130] Based on the output of the spatiotemporal convolutional network, tornado prediction control variables are generated, including the wind speed field. Pressure field p and temperature field T c And represent it as a four-dimensional tensor with time, height, width, and channel;

[0131] The Navier-Stokes equation, continuity equation and thermodynamic energy equation are used as physical constraints to ensure that the tornado prediction results meet the laws of fluid mechanics and thermodynamics; the three equations are as follows:

[0132] ,

[0133] In the formula, is a partial derivative, t is time, is a wind speed vector, T c is temperature, is air pressure, is density, is a kinematic viscosity coefficient, is a thermal diffusion coefficient, is an external force term, is a heat source term, is a gradient operator;

[0134] 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 :

[0135] ,

[0136] In the formula, F is the residual function of the physical equation, is the prediction point from 1 to N summed up;

[0137] The prediction control quantity is compared with the Doppler weather radar and ECMWF observation data, and the root mean square error :

[0138] ,

[0139] In the formula, is the prediction value, is the observation value, is the observation point from 1 to N summed up;

[0140] The Dirichlet boundary condition is applied to the prediction area, and the deviation of the difference between the boundary point prediction value and the boundary condition is calculated :

[0141] ,

[0142] In the formula, is the boundary prediction value, is a boundary value, is a boundary point from 1 to N iterative summation;

[0143] joint total loss function is defined as:

[0144] ,

[0145] wherein, are weight coefficients respectively; based on the total loss function , the deep neural network parameters are optimized to complete model training.

[0146] In step 5, the output probability quantization step specifically includes:

[0147] According to the meteorological forecasting requirements, the prediction time window is set to be 1 hour in the future, and the time resolution is 5 minutes; as shown in Figure 4 A probability prediction module is added after the network output layer, using an activation function to map the predicted control amount at each time point to the interval [0, 1], generating the tornado occurrence probability of each grid point, forming a four-dimensional probability tensor of time, height, width, and probability value;

[0148] Based on historical data, the receiver operating characteristic curve and performance chart of the probability distribution are calculated, and the threshold value is optimized to maximize the critical success index;

[0149] By applying the threshold value, a binary mask is generated to mark the potential tornado generation area; a connected domain analysis algorithm is used to cluster high-probability areas to generate continuous potential generation area boundaries; Gaussian smoothing filtering is applied to preserve the main high-probability area morphology;

[0150] 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 map in the form of error bars; based on the probability threshold and the confidence, the low, medium, and high risk warning levels are divided;

[0151] Finally, the tornado time series prediction probability map is generated, outputting a two-dimensional slice according to the ground height, marking the high-probability area boundary, confidence interval, and warning level, providing intuitive warning information for tornado forecasters.

Claims

1. A tornado spatiotemporal intelligent prediction method fusing multi-source data and physical information, characterized in that, The method comprises the following steps: Step 1, multi-source data collection: collect the occurrence time, location and intensity of different tornado events, and collect the corresponding Doppler weather radar observation data and ECMWF reanalysis data of the medium-term weather forecasting center; 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 a high-resolution Cartesian grid, 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 a spatio-temporal convolutional network: input the data samples into a spatio-temporal convolutional network composed of an encoder, a jump connection and a decoder, and finally output a tornado prediction control quantity; The encoder has a nested architecture of a spatio-temporal convolutional layer composed of a convolutional long short-term memory network unit, an activation function and a maximum pooling downsampling, a coordinate convolutional layer and a large kernel selection convolutional module, which are used to capture 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 gradually restores the spatial resolution of the feature map based on a bilinear upsampling considering coordinate information and a convolutional long short-term memory network unit; 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 a joint loss function including a control equation loss function, an observation data loss function and a boundary condition loss function is defined, and the spatio-temporal convolutional network is optimized based on the joint loss function to predict the future tornado; specifically including: According to the spatiotemporal convolution network output, a tornado prediction control quantity is generated, including a wind speed field (u, v, w), a pressure field p and a 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, the continuity equation and the thermodynamic energy equation are used as physical constraints to ensure that the tornado prediction result meets the laws of fluid mechanics and thermodynamics; the forms of the three equations are as follows: wherein is the partial derivative, t is the time step, u is the wind velocity vector, T c is the temperature, p is the pressure, p is the density, v is the kinematic viscosity, k is the thermal diffusivity, f is the external force term, Q is the heat source term, is the gradient operator; The partial derivatives of the predicted control amount with respect to time and space are 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 L is calculated phys : where F(·) is a residual function of the physical equation, is a summation over the iterations from 1 to N of the prediction point i; The forecast control amounts are compared with the weather radar and ECMWF data, and their root mean square errors L are calculated data : where y pred is the predicted value, y obs is the observed value, is the summation over the observations i from 1 to N. Applying Dirichlet boundary conditions to the prediction region, computing a bias L of the difference between the predicted value of the boundary points and the boundary conditions bc : where y pred_bc is the boundary prediction value, y bc is the boundary value, is the sum over the iteration of the boundary point i from 1 to N. Joint total loss function L total is defined as: L total = λ1L phys + λ2L data + λ3L bc (22) In the formula, λ1, λ2, λ3 are weight coefficients respectively; the back propagation algorithm is adopted to optimize the deep neural network parameters based on the total loss function L total , and model training is completed. Step 5, output probability quantization: set different time windows for the spatio-temporal convolutional network, obtain the tornado prediction control quantity at different time points 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 the confidence interval, thereby providing clear early warning decision support for meteorological forecast personnel; specifically including: According to the meteorological forecasting requirements, the prediction time window and the time resolution are set; a probability prediction module is added after the network output layer, and an activation function is used to scale the prediction control quantity at each time point to the [0, 1] interval, thereby generating the tornado occurrence probability of 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 the performance diagram of the probability distribution are calculated, and the threshold is optimized to maximize the key success index; A binary mask is generated by applying the threshold, and the potential tornado generation area is marked; a connected component analysis algorithm is used to cluster the high-probability area to generate a continuous potential generation area boundary; a Gaussian smoothing filter is applied 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.

2. The tornado spatio-temporal intelligent prediction method of fusing multi-source data with physical information according to claim 1, characterized in that, In step 1, the multi-source data collection step specifically includes: Statistical analysis of historical tornado event records in the national meteorological disaster database, structured storage of important information such as start and end time, latitude and longitude coordinates, enhanced Fujita scale, path length and maximum width, and disaster situation of each tornado event; According to the central longitude and latitude (φ T ,λ T ) of each tornado event, the nearest next-generation weather radar site (φ R ,λ R ) is queried and determined; the distance d between the two is calculated to ensure that it is within the effective detection radius; the distance d calculation formula is as follows: where r is the earth radius, Δφ = φ T -φ R is the latitude difference, Δλ = λ T -λ R is the longitude difference, sin(·) and cos(·) are the sine and cosine functions, respectively, and arcsin(·) is the inverse sine function. Define a one-hour data collection time window before and after each tornado event; within this time and space range, download all available raw radar-based data files from the Level-II data archive, including dual-polarization products such as reflectivity factor DBZ, radial velocity VEL, spectral width WIDTH, specific differential phase shift KDP, correlation coefficient RHOHV, and differential reflectivity ZDR; Take the center latitude and longitude of each tornado event as the geometric center, define a rectangular area with latitude and longitude as the range; use the same time window as the radar data, download the ECMWF reanalysis data within this time period, including physical quantities such as 2-meter height air temperature, 2-meter height dew point temperature, 10-meter height east-west and north-south wind components, sea level pressure, and standard isobaric surface layer potential height, temperature, east-west / north-south wind components, and relative humidity / specific humidity at 250, 300, 500, 700, and 850 hPa.

3. The tornado spatio-temporal intelligent prediction method of fusing multi-source data with physical information according to claim 2, characterized in that, In step 2, feature extraction and fusion specifically includes: Use Lambert conformal conic projection to set the Cartesian grid center point as the starting latitude and longitude coordinates of each tornado event; set the grid coverage range and grid resolution to finally 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 one hour of continuous observation before and after the tornado occurs; For each original radar-based data file, quality control is performed, including: filtering of ground clutter and non-meteorological echoes, de-aliasing of radial velocity; the quality-controlled polar coordinate radar data is interpolated to the aforementioned defined Cartesian grid by the Barnes objective analysis method, and the grid point values are estimated by distance-weighted averaging of neighboring observation points, with the core weight function w i and the calculation of grid interpolation G is as follows: where N is the number of observation points, exp(·) is the natural exponential function; O i and w i are the value and weight of the i-th radar observation point, respectively, d i is the distance of each point to the i-th observation point, r and l0are adjustable parameters that affect the degree of smoothing and the scale of response, is the summation over the observation points i from 1 to N. Select three low, medium and high altitudes to generate constant height position plan; at each height, extract reflectivity factor DBZ, radial velocity VEL, spectral width WIDTH, specific differential phase shift KDP, correlation coefficient RHOHV and differential reflectivity ZDR radar physical quantities as independent feature channels; In different resolution ECMWF reanalysis data, bilinear interpolation method is used for spatial low resolution ECMWF reanalysis data to upsample to high resolution Cartesian grid consistent with radar data; linear time interpolation is used to generate environment field base data consistent with radar data step by step from hourly resolution reanalysis data; on the interpolated high resolution grid, calculate the key environmental parameters of 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 g is the acceleration of gravity, z is the specified height, z LFC and z EL are the free convection height and the equilibrium height, respectively, T vp is the virtual temperature of the test air, T ve is the virtual temperature of the ambient air; z1 and z2 are the heights of the lower and upper layers, respectively; u(z1) and u(z2) are the east-west wind speed components at heights z1 and z2, respectively, and v(z1) and v(z2) are the north-south wind speed components at heights z1 and z2, respectively; is the horizontal wind vector at height z, is the estimated storm movement vector, and ω(z) is the horizontal vorticity vector. For each time step, the feature maps generated by all the radar height layers and ECMWF environmental parameters are stacked along the channel dimension; the three-dimensional height, width, and channel feature maps of multiple time steps, stacked channel dimension, are stacked again in time order, finally forming a four-dimensional 4D tensor sample X with a dimension of [time step, height, width, channel number], and its structure can be represented as: In the formula, is a four-dimensional array, T is the number of continuous time steps, H is the grid height, W is the grid width, and C is the total number of feature channels.

4. The tornado spatio-temporal intelligent prediction method of fusing multi-source data with physical information according to claim 3, characterized in that, In step 3, the tornado prediction based on the spatio-temporal convolutional network specifically includes: 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 normalized fusion feature data samples 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 down-sampling operations are performed step by step; different layers are set with different numbers of convolution kernels for hierarchical feature extraction and dimension transformation; 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 unit can extract time series and radar spatial features at the same time, and the calculation formula is as follows: where subscript t is the 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, i t , f t , o t are the input gate, forget gate, and output gate, respectively, W xi , W hi , W ci are the convolution kernel weights for the input gate corresponding operation, W xf , W hf , W cf are the convolution kernel weights for the forget gate corresponding operation, W xc , W hc are the convolution kernel weights for the cell state corresponding operation, W xo , W ho , W co are the convolution kernel weights for the output gate corresponding operation, b i , b f , b c , b o are the bias terms, and * is the convolution operation, is the element-wise multiplication, and σ and tanh are the activation functions. Then, the coordinate convolution layer with coordinate information is embedded into the input tensor of the spatio-temporal convolutional layer to provide additional coordinate system spatial layout; the kernel size of the convolution kernel in the spatio-temporal convolutional module is dynamically adjusted through the large kernel selection convolution module to handle the differences in context information between different tornado targets, and the calculation formula is as follows: where Y is the output variable, X is the input variable, F(·) is the convolution operation, S(·) is the activation function, P avg (·) is the average pooling, P max (·) is the max pooling operation, F i 1×1×1 (·) is the 1x1x1 convolution, F 2→N converts the 2-channel pooled features into N spatial attention feature maps, U i is the i-th layer feature of X, N is the number of convolution kernels, is the traversal summation of feature map i from 1 to N; 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, σ1, σ2 are activation functions; x l , g are features of the decoder and the encoder respectively, ψ is a linear change operation after σ1 activation, superscript l is a linear mapping dimension space, superscript T is a transpose operation; and ψ T are linear change transposes after x l , g and σ1 activation respectively; b ψ , b g are bias terms of ψ and g respectively; is an attention score, Θ att is a feature parameter, is an attention coefficient; A decoder considering the bilinear up-sampling of coordinate information and two spatio-temporal convolutional layers is established to double the spatial dimension of the tornado spatio-temporal feature map combined with 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.

Images