Heavy rain prediction method, device, equipment and medium based on four-dimensional space-time perception

Abstract

This application discloses a rainstorm prediction method, apparatus, device, and medium based on four-dimensional spatiotemporal perception. The method includes: generating a four-dimensional tensor including interrelated time series, vertical layers, spatial grids, and meteorological variables; performing spatial feature extraction and temporal modeling on the four-dimensional tensor based on a target 3D convolution model obtained by training the four-dimensional tensor and a joint constraint function to obtain a first feature tensor of spatiotemporal union; determining the rainstorm risk weight at the spatial position corresponding to each spatial grid, fusing the weight with the first feature tensor to obtain a second feature tensor, and performing a rolling decoding process based on the time series on the second feature tensor to obtain a rainstorm prediction result. This application fully extracts the local features of the four-dimensional tensor in the vertical layers and horizontal space of the atmosphere through a 3D convolution model, and decodes the tensor in the time dimension to obtain a prediction result, thereby achieving spatiotemporal perception of meteorological variables and improving the accuracy of rainstorm prediction results.

Description

Technical Field

[0001] The present application relates to, but is not limited to, the field of heavy rain prediction technology, and in particular to a heavy rain prediction method, device, equipment, and medium based on four-dimensional space-time perception. Background Art

[0002] With the convergence of artificial intelligence and meteorological science, deep learning has demonstrated its powerful capabilities in predicting short- to medium-term severe convective weather, thunderstorms, heavy rain, and other extreme events. For example, the Transformer architecture excels at time series modeling, while convolutional neural networks excel at extracting spatial features, driving the development of various hybrid forecasting models. However, existing deep learning-based heavy rain forecasting methods suffer from inadequate time series modeling, neglect of the continuity of precipitation evolution, and weak modeling of meteorological data at different atmospheric altitudes or pressure levels, failing to reflect the interactions between high- and low-altitude systems. Consequently, existing heavy rain forecasts lack accuracy. Summary of the Invention

[0003] The embodiments of the present application provide a rainstorm prediction method, apparatus, device and medium based on four-dimensional space-time perception, which can effectively improve the accuracy of rainstorm prediction results.

[0004] In a first aspect, an embodiment of the present application provides a rainstorm prediction method based on four-dimensional space-time perception, comprising:

[0005] Acquire multiple initial meteorological variable parameters, and generate a four-dimensional tensor based on the initial meteorological variable parameters, wherein the four-dimensional tensor includes interrelated time series, vertical layers, spatial grids, and meteorological variables, and the vertical layers are the number of vertical atmospheric layers;

[0006] Performing spatial feature extraction and temporal modeling on the four-dimensional tensor based on a trained target 3D convolution model to obtain a first spatiotemporal joint feature tensor, wherein the target 3D convolution model is trained based on the four-dimensional tensor and a joint constraint function, wherein the joint constraint function is composed of a water vapor conservation loss function, an energy conservation loss function, and a spatial gradient regularization constraint function;

[0007] Determining a rainstorm risk weight at a spatial location corresponding to each of the spatial grids based on a historical rainstorm event dataset;

[0008] Performing feature fusion on the rainstorm risk weight and the first feature tensor to obtain a second feature tensor with enhanced features;

[0009] The second feature tensor is subjected to rolling decoding processing based on a time series to obtain a spatiotemporally continuous heavy rain prediction result.

[0010] In some embodiments, generating a four-dimensional tensor based on the initial meteorological variable parameters includes:

[0011] Resampling all of the initial meteorological variable parameters to a target spatial grid using a bilinear interpolation method to obtain a first intermediate variable;

[0012] Unifying the original time resolutions of all the first intermediate variables to a target time frequency using a linear time interpolation method to obtain a second intermediate variable;

[0013] Performing data correction and data cleaning on all the second intermediate variables to obtain third intermediate variables;

[0014] Adjusting all the third intermediate variables into respective fourth intermediate variables, wherein the structured data format of the spatial grids corresponding to the respective fourth intermediate variables is the same;

[0015] Based on each time step, concatenating the fourth intermediate variables at each vertical layer and each spatial grid to obtain a multi-channel tensor, wherein the time step belongs to the time series;

[0016] The multi-channel tensors are stacked according to the time dimension to obtain the four-dimensional tensor.

[0017] In some embodiments, spatial feature extraction and temporal modeling are performed on the four-dimensional tensor based on the trained target 3D convolutional model to obtain a first spatial-temporal joint feature tensor, including:

[0018] For each meteorological variable in the four-dimensional tensor, based on each time step of the time series, the meteorological variable is mapped into a first intermediate tensor of a unified embedding dimension through a first convolutional layer or a linear layer;

[0019] Input the intermediate tensor into the target 3D convolutional model, and generate a query Q vector, a key K vector, and a value V vector based on the first intermediate tensor at each time step and spatial position;

[0020] Calculate the attention weight based on the Q vector and K vector at each time step;

[0021] Performing weighted aggregation on the V vectors based on the respective attention weights to obtain respective second intermediate tensors;

[0022] Stacking each of the second intermediate tensors according to the time dimension to obtain a third intermediate tensor;

[0023] Perform residual processing on the third intermediate tensor to obtain the first feature tensor.

[0024] In some embodiments, the historical rainstorm event dataset includes rainstorm events in a preset year, and determining the rainstorm risk weight at the spatial location corresponding to each spatial grid based on the historical rainstorm event dataset includes:

[0025] Counting the occurrence frequency of the rainstorm event on each spatial grid and forming a frequency heat map;

[0026] The frequency heat map is normalized into a significance probability map, wherein the significance probability map is used to indicate the rainstorm risk weight of the spatial position corresponding to each of the spatial grids.

[0027] In some embodiments, the rainstorm risk weight is subjected to feature fusion with the first feature tensor to obtain a second feature tensor after feature enhancement, including:

[0028] Expanding the dimension of the saliency probability map based on the first feature tensor to obtain a reference saliency probability map, wherein the number of dimensions of the first feature tensor is the same as the number of dimensions of the reference saliency probability map;

[0029] The reference saliency probability map is weightedly fused with the first feature tensor pixel by pixel to obtain the second feature tensor.

[0030] In some embodiments, the second feature tensor is subjected to a rolling decoding process based on a time series to obtain a spatiotemporally continuous rainstorm prediction result, including:

[0031] Performing spatial dimension enhancement processing on the second feature tensor at each time step through the upsampling structure and convolution structure of the decoder to obtain a third feature tensor, wherein the spatial dimension of the third feature tensor is the same as the spatial dimension of the four-dimensional tensor;

[0032] Compressing the meteorological variable dimension of each of the third feature tensors to 1 through a second convolutional layer to obtain each fourth feature tensor;

[0033] All of the fourth feature tensors are spliced ​​in chronological order to obtain the rainstorm prediction result.

[0034] In some embodiments, the target 3D convolutional model is trained according to the following steps:

[0035] Inputting the four-dimensional tensor into an initial 3D convolution model and outputting an initial rainstorm prediction result;

[0036] Calculating a loss value of the initial 3D convolutional model based on the joint constraint function and combining the initial rainstorm prediction result and the corresponding true prediction result;

[0037] When the loss value is greater than a preset loss threshold, the model parameters of the initial 3D convolution model are iteratively adjusted based on the loss value and the loss threshold until the loss value is less than or equal to the loss threshold, thereby obtaining the target 3D convolution model.

[0038] In a second aspect, an embodiment of the present application provides a control device comprising at least one control processor and a memory for communicating with the at least one control processor; the memory stores instructions that can be executed by the at least one control processor, and the instructions are executed by the at least one control processor so that the at least one control processor can execute the heavy rain prediction method based on four-dimensional space-time perception as described in the first aspect.

[0039] In a third aspect, an embodiment of the present application further provides an electronic device comprising the control device of the second aspect.

[0040] In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are used to execute the heavy rain prediction method based on four-dimensional space-time perception as described in the first aspect.

[0041] An embodiment of the present application provides a rainstorm prediction method, apparatus, device and medium based on four-dimensional space-time perception, the method comprising: obtaining a plurality of initial meteorological variable parameters, generating a four-dimensional tensor based on the initial meteorological variable parameters, wherein the four-dimensional tensor comprises interrelated time series, vertical layers, spatial grids and meteorological variables, and the vertical layers are the number of vertical atmospheric layers; performing spatial feature extraction and temporal modeling on the four-dimensional tensor based on a trained target 3D convolution model to obtain a first feature tensor of space-time joint, wherein the target 3D convolution model is trained based on the four-dimensional tensor and a joint constraint function, and the joint constraint function is composed of a water vapor conservation loss function, an energy conservation loss function and a spatial gradient regularization constraint function; determining the rainstorm risk weight at the spatial position corresponding to each of the spatial grids based on a historical rainstorm event data set; performing feature fusion on the rainstorm risk weight and the first feature tensor to obtain a second feature tensor with enhanced features; performing rolling decoding processing on the second feature tensor based on a time series to obtain a spatiotemporal continuous rainstorm prediction result. According to the solution provided in the embodiment of the present application, the local features of the four-dimensional tensor in the vertical layer and horizontal space of the atmosphere are fully extracted through the 3D convolution model, and the local features are decoded and processed in the time dimension, paying more attention to the temporal continuity of precipitation evolution, and fully considering the changes in meteorological variables at different altitudes and different pressure layers in the atmosphere, that is, realizing the spatiotemporal perception of meteorological variables. Compared with the existing methods, the heavy rain prediction results obtained by the solution of the present application are more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a flowchart of the steps of a rainstorm prediction method based on four-dimensional space-time perception provided by an embodiment of the present application;

[0043] Figure 2 This is a structural diagram of a control device provided in another embodiment of the present application. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0045] It is understood that although the device schematics illustrate functional module divisions and the flowcharts illustrate logical sequences, in certain circumstances, the steps shown or described may be performed in a sequence that differs from the module divisions in the device or the sequence in the flowcharts. The terms "first," "second," and the like in the specification, claims, or accompanying drawings are used to distinguish similar items and are not necessarily used to describe a specific sequence or precedence.

[0046] With the convergence of artificial intelligence and meteorological science, deep learning has demonstrated its powerful capabilities in predicting short- to medium-term severe convective weather, thunderstorms, heavy rain, and other extreme events. For example, the Transformer architecture excels at time series modeling, while convolutional neural networks excel at extracting spatial features, driving the development of various hybrid forecasting models. However, existing deep learning-based heavy rain forecasting methods suffer from inadequate time series modeling, neglect of the continuity of precipitation evolution, and weak modeling of meteorological data at different atmospheric altitudes or pressure levels, failing to reflect the interactions between high- and low-altitude systems. Consequently, existing heavy rain forecasts lack accuracy.

[0047] To solve the above-mentioned problems, an embodiment of the present application provides a rainstorm prediction method, device, equipment and medium based on four-dimensional space-time perception, the method comprising: obtaining multiple initial meteorological variable parameters, generating a four-dimensional tensor based on the initial meteorological variable parameters, wherein the four-dimensional tensor includes interrelated time series, vertical layers, spatial grids and meteorological variables, and the vertical layers are the number of vertical atmospheric layers; performing spatial feature extraction and temporal modeling on the four-dimensional tensor based on a trained target 3D convolution model to obtain a first feature tensor of space-time joint, wherein the target 3D convolution model is trained based on the four-dimensional tensor and a joint constraint function, and the joint constraint function is composed of a water vapor conservation loss function, an energy conservation loss function and a spatial gradient regularization constraint function; determining the rainstorm risk weight at the spatial position corresponding to each of the spatial grids based on a historical rainstorm event data set; performing feature fusion on the rainstorm risk weight and the first feature tensor to obtain a second feature tensor with enhanced features; performing rolling decoding processing on the second feature tensor based on a time series to obtain a spatiotemporal continuous rainstorm prediction result. According to the solution provided in the embodiment of the present application, the local features of the four-dimensional tensor in the vertical layer and horizontal space of the atmosphere are fully extracted through the 3D convolution model, and the local features are decoded and processed in the time dimension, paying more attention to the temporal continuity of precipitation evolution, and fully considering the changes in meteorological variables at different altitudes and different pressure layers in the atmosphere, that is, realizing the spatiotemporal perception of meteorological variables. Compared with the existing methods, the heavy rain prediction results obtained by the solution of the present application are more accurate.

[0048] The embodiments of the present application are further described below with reference to the accompanying drawings.

[0049] refer to Figure 1 , Figure 1 This is a flowchart of a rainstorm prediction method based on four-dimensional space-time perception provided by an embodiment of the present application. The embodiment of the present application provides a rainstorm prediction method based on four-dimensional space-time perception, which includes but is not limited to the following steps:

[0050] Step S10, obtaining multiple initial meteorological variable parameters, and generating a four-dimensional tensor based on the initial meteorological variable parameters, wherein the four-dimensional tensor includes interrelated time series, vertical layers, spatial grids and meteorological variables, and the vertical layers are the number of vertical atmospheric layers.

[0051] Specifically, the initial meteorological variable parameters in this embodiment are collected from multiple sources, including ECMWF / ERA5, FY-4, ground station observations, and radar echoes. The collected initial meteorological variable parameters include ECMWF / ERA5 reanalysis data, FY-4 satellite products, ground-based automatic weather station observations, and weather radar echoes. Reanalysis data provides hourly atmospheric variables (such as specific humidity, temperature, wind speed, and air pressure), with a three-dimensional structure covering multiple layers of pressure surfaces. Satellite remote sensing data provides two-dimensional observations such as cloud images, brightness temperature, and precipitation estimates. Ground stations provide hourly measured air temperature, precipitation, wind speed, and other data. Weather radar echo data provides near-surface convective reflectivity with a resolution of up to 0.01° × 0.01° and a time interval of 5-6 minutes. It can be understood that acquiring these multi-source parameters provides an effective data foundation for subsequently obtaining accurate heavy rain forecasts.

[0052] Specifically, in some embodiments, Figure 1 Step S10 includes but is not limited to the following steps:

[0053] Step S11, using bilinear interpolation method to resample all initial meteorological variable parameters to the target spatial grid to obtain the first intermediate variable;

[0054] Step S12: using a linear time interpolation method to unify the original time resolutions of all first intermediate variables to a target time frequency, thereby obtaining a second intermediate variable;

[0055] Step S13, performing data correction and data cleaning on all second intermediate variables to obtain third intermediate variables;

[0056] Step S14: adjusting all the third intermediate variables into respective fourth intermediate variables, wherein the structured data format of the spatial grids corresponding to the respective fourth intermediate variables is the same;

[0057] Step S15: Based on each time step, concatenate the fourth intermediate variables of each vertical layer and each spatial grid to obtain a multi-channel tensor, wherein the time step belongs to the time series;

[0058] Step S16: stack the multi-channel tensors according to the time dimension to obtain a four-dimensional tensor.

[0059] As can be understood from the description of the above embodiment, the initial meteorological variable parameters from multiple sources are heterogeneous in terms of spatial resolution, temporal granularity, and physical quantity type, requiring unified processing. This embodiment ensures that the spatial structures corresponding to the variable parameters are unified to facilitate the subsequent use of spatial convolution to extract structured features and avoid spatial mismatches.

[0060] The specific steps include: (1) using bilinear interpolation to resample all initial meteorological variable parameters to the target spatial grid (0.25° × 0.25° in this embodiment) to obtain the first intermediate variable, where the corresponding interpolation formula is as follows:

[0061] ;

[0062] in, are the values ​​of the four adjacent grid points in the initial meteorological variable parameters, is the linear weight calculated by distance, satisfy , are the grid coordinates of the target space grid.

[0063] It is understandable that the use of bilinear interpolation for structured grids (such as the regular latitude and longitude grids commonly used in meteorological data) is highly efficient, can ensure continuity and smoothness, and is suitable for large-scale grid data; it is conducive to subsequent rapid deployment in large-scale meteorological systems; in addition, the determinism of the interpolation structure ensures the subsequent model's stable perception of the spatial structure of the input data, which is one of the foundations for the establishment of physical consistency constraints.

[0064] It is understandable that this embodiment uses the linear time interpolation method to unify the original time resolutions of all first intermediate variables (for example, scattered 6 minutes, 15 minutes, 1 hour, etc.) into a target time frequency (such as 1 hour) to obtain the second intermediate variable. For the original time resolution of less than 1 hour, this embodiment uses the linear time interpolation method to perform linear interpolation on the missing time points to ensure that there is a complete variable value every hour. The corresponding expression of the linear time interpolation method is as follows:

[0065] ;

[0066] in, is the target time step (moment) t The variable value (that needs to be interpolated), is the closest known historical time point to 𝑡, is the closest known future time point to 𝑡, It's time The known value at t∈ It can be understood that the expression corresponding to the linear time interpolation method is a one-variable linear function in the interval [ If I know the value of a variable at two time points (such as 10:00 and 11:00), I can use these two values ​​to linearly estimate the value at an intermediate time (such as 10:30) in proportion to the time. In smaller time scales, the change of the variable is linear.

[0067] In step S20, spatial feature extraction and temporal modeling are performed on the four-dimensional tensor based on the trained target 3D convolution model to obtain a first spatial-temporal joint feature tensor, wherein the target 3D convolution model is trained based on the four-dimensional tensor and the joint constraint function, and the joint constraint function is composed of a water vapor conservation loss function, an energy conservation loss function, and a spatial gradient regularization constraint function.

[0068] For example: Assume: The temperature at 10:00 The temperature at 11:00 is 20℃ The temperature is 22℃. To interpolate the temperature at 10:30, t=10:30. Substitute into the above formula:

[0069] ;

[0070] It can be seen that from 10:00 to 10:30 to 11:00, the temperature changes linearly.

[0071] Next, this embodiment performs data correction and data cleaning on all the second intermediate variables to obtain the third intermediate variable. This embodiment does not limit the data correction and data cleaning methods according to the graph, and can be normalization processing such as unit conversion, anomaly elimination, missing value filling, terrain correction, etc. In this way, the cleaned data provides physically meaningful and stable input for subsequent tensor generation and modeling, which can improve the stability of model training and is the prerequisite for constructing a physical consistency loss term.

[0072] Next, based on each time step, the third intermediate variables of each vertical layer and each spatial grid are spliced ​​to obtain a multi-channel tensor, where the time step belongs to the time series. The multi-channel tensors are stacked according to the time dimension to obtain a four-dimensional tensor. In this way, all the initial meteorological variable parameters can be output as a structured data format of a unified spatial grid in the following form:

[0073] ;

[0074] Where T represents the time series (i.e. the number of time steps, such as 24 hours), L represents the number of vertical layers, H × W are the grid points of the spatial grid, F is a meteorological variable, X is the fourth intermediate variable.

[0075] Furthermore, this embodiment splices the fourth intermediate variables under each vertical layer and each spatial grid based on each time step to obtain a multi-channel tensor, wherein the time step belongs to the time series, and the multi-channel tensor is stacked according to the time dimension to obtain a four-dimensional tensor. That is, the pre-processed multi-source meteorological observation variables are structured into a unified high-dimensional tensor input for deep learning models to use for spatiotemporal modeling. Taking into account that the rainstorm process has significant "vertical-space-time-multivariable" coupling characteristics, this embodiment adopts the "time × layer × space × channel" four-dimensional tensor to form a complete meteorological input data tensor. Specifically, the fourth intermediate variable includes the time step t , vertical layers L (such as the number of layers corresponding to 500hPa, 700hPa, and 850hPa), spatial grid H × W , meteorological variables F (such as temperature, specific humidity, U wind, V wind, geopotential height). Based on each time step t , the vertical layers L And the fourth intermediate variable under each spatial grid is spliced ​​to obtain a multi-channel tensor, where the time step belongs to the time series T (T contains multiple t), the multi-channel tensor The expression is:

[0076] ;

[0077] in, They are temperature, specific humidity, U wind, V wind and geopotential height, at different vertical layers or different spatial grids. Meteorological variables have corresponding hierarchical distinctions.

[0078] In this embodiment, the multi-channel tensors are stacked according to the time dimension to obtain a four-dimensional tensor The expression is as follows:

[0079] .

[0080] Specifically, in some embodiments, Figure 1 Step S20 includes but is not limited to the following steps:

[0081] Step S21: for each meteorological variable in the four-dimensional tensor, based on each time step of the time series, the meteorological variable is mapped into a first intermediate tensor of a unified embedding dimension through a first convolutional layer or a linear layer;

[0082] Step S22: input the intermediate tensor into the target 3D convolutional model, and generate a query Q vector, a key K vector, and a value V vector based on the first intermediate tensor at each time step and spatial position;

[0083] Step S23, calculating the attention weight based on the Q vector and K vector of each time step;

[0084] Step S24, performing weighted aggregation on the V vectors based on the respective attention weights to obtain respective second intermediate tensors;

[0085] Step S25: stacking the second intermediate tensors according to the time dimension to obtain a third intermediate tensor;

[0086] Step S26: Perform residual processing on the third intermediate tensor to obtain a first feature tensor.

[0087] It can be understood that the purpose of using the target 3D convolutional model in this embodiment is to establish a joint representation (i.e., the first feature tensor) of the input variable (i.e., the four-dimensional tensor) at multiple vertical levels (L), multiple spatial grid scales (H×W), and multiple time steps (T) through the dual mechanisms of spatial feature extraction and temporal dynamic modeling.

[0088] It should be noted that in this embodiment, each meteorological variable in the four-dimensional tensor is based on each time step of the time series. t , the meteorological variables are mapped into a first intermediate tensor of uniform embedding dimension through the first convolutional layer or linear layer. The first intermediate tensor is calculated according to the following formula:

[0089] ;

[0090] in, Indicates the t The data tensor of time steps, the original shape is , used to input the first convolutional layer, Conv1x1, where F is the meteorological variable (such as temperature, humidity, wind speed, etc., generally 5-10). Conv1x1 uses a three-dimensional convolution kernel of size 1×1×1 to complete the linear mapping of the channel dimension (equivalent to point-by-point full connection). C is the channel embedding dimension (can be a hyperparameter, such as 32, 64, etc.), indicating the feature dimension after compression or expansion. It is the intermediate tensor after embedding, with a shape of C×L×H×W, which serves as the input of the target 3D convolutional model.

[0091] It's understandable that step S21 primarily fuses the features of different meteorological variables (temperature, humidity, wind speed, etc.). This is equivalent to learning a linear transformation matrix to map the original variable space into a more compact representation space. Nonlinear activations (such as ReLU) can also be introduced to further enhance feature representation. This process achieves both physical variable feature fusion and dimensionality regularization.

[0092] Furthermore, the intermediate tensor Input to the target 3D convolutional model, generate query Q vector, key K vector and value V vector based on the first intermediate tensor of each time step and spatial position, specifically, each Input multi-layer 3D convolutional neural network, L × H × W Slide upward to extract spatial-hierarchical joint local features, and the feature extraction range covers the high-low altitude interaction and adjacent spatial grids (such as frontal areas and low-pressure centers). The expression for constructing spatial features is as follows:

[0093] ;

[0094] in, It is t The spatial feature representation vector of time steps (i.e., the first intermediate tensor) has the shape , Represents a three-dimensional convolution operation (including multiple convolution layers), They represent the vertical number of layers, height, and width of the output tensor (determined by the input size, convolution kernel size, stride, and padding). D is the number of output channels, i.e., the number of dimensions of the extracted spatial structure features. The convolution kernel weight parameters are automatically learned through backpropagation during model training and do not require manual specification.

[0095] Furthermore, for each spatial position ( L , H,W ) on the first intermediate tensor sequence [ ,…, ] for temporal modeling. Construct query Q vector, key K vector and value V vector, perform cross-attention calculation, and output the spatiotemporal joint representation at each moment. That is, calculate the attention weight based on the Q vector and K vector of each time step, perform weighted aggregation on the V vector based on each attention weight to obtain each second intermediate tensor, and stack each second intermediate tensor according to the time dimension to obtain the third intermediate tensor.

[0096] It should be noted that the attention weight is calculated based on the Q vector and K vector of each time step, and is calculated according to the following formula:

[0097] ;

[0098] It should be noted that the V vector is weightedly aggregated based on each attention weight to obtain each second intermediate tensor, and each second intermediate tensor is stacked according to the time dimension to obtain the third intermediate tensor, which is calculated according to the following formula:

[0099] ;

[0100] in, 、 and Respectively indicate time and Q vector, K vector and value V, Indicates time and The attention weight of the correlation between is the third intermediate tensor.

[0101] Furthermore, this embodiment introduces a residual mechanism to enhance information flow and gradient propagation. This involves performing residual processing on the third intermediate tensor to obtain a first feature tensor. This prevents deep network degradation and improves modeling stability and generalization capabilities. In this embodiment, after step S26, the expression for outputting the first feature tensor is as follows:

[0102] ;

[0103] in, It is the first feature tensor, which embeds the local spatial structure and dynamic time evolution information for further enhancement by the subsequent saliency guidance module.

[0104] Specifically, in some embodiments, the target 3D convolutional model is trained according to the following steps:

[0105] Step S27, inputting the four-dimensional tensor into the initial 3D convolution model and outputting the initial heavy rain prediction result;

[0106] Step S28, based on the joint constraint function, the loss value of the initial 3D convolutional model is calculated by combining the initial rainstorm prediction result and the corresponding true prediction result;

[0107] In step S29, when the loss value is greater than the preset loss threshold, the model parameters of the initial 3D convolution model are iteratively adjusted based on the loss value and the loss threshold until the loss value is less than or equal to the loss threshold, thereby obtaining the target 3D convolution model.

[0108] It should be noted that the expression of the joint constraint function in this embodiment is as follows:

[0109] ;

[0110] in, for, is the mean square error term, is the water vapor conservation loss function, is the energy conservation loss function, is the spatial gradient regularization constraint function, 、 and is the loss weighting coefficient, which can be tuned through the validation set. 、 、 These three constraints form a complementary conservation and regularization system, which synergistically improve the model's performance in "realism + smoothness + dynamic consistency".

[0111] Among them, the constraint principle of the water vapor conservation loss function is: the formation of precipitation in the atmosphere must come from changes in water vapor flux. If the precipitation prediction map is seriously inconsistent with the changes in the input water vapor flow field, it means that the prediction result lacks physical consistency. The expression is as follows:

[0112] ;

[0113] in, and They represent the water vapor flux in the predicted results and the corresponding real water vapor flux, which are calculated from wind speed and specific humidity. is a divergence operator used to calculate the divergence of water vapor flux. This formula can measure the degree of deviation between the predicted field and the input field in terms of water vapor conservation, providing effective support for avoiding the generation of "false precipitation" by the model and improving the physical reliability of the forecast.

[0114] Among them, the constraint principle of the energy conservation loss function is: the total energy (kinetic energy + potential energy + internal energy) in the meteorological system is approximately conserved on a short time scale. If the model prediction output causes a drastic change in the total energy, it indicates that the internal mechanism is unbalanced. The expression is as follows:

[0115] ;

[0116] in, and They represent the total energy in the prediction results and the corresponding true total energy, respectively. This formula can measure the degree of deviation between the prediction field and the input field in terms of energy conservation, constrain the model to maintain energy balance, and avoid non-realistic violent disturbances (such as local "explosive" rainfall).

[0117] In addition, the total energy E The calculation formula is as follows:

[0118] ;

[0119] in, u and v is the horizontal wind speed, g is the gravitational constant, z is the potential height, is the specific heat capacity, T 1 is the temperature.

[0120] Among them, the constraint principle of the spatial gradient regularization constraint function is: the actual meteorological field is continuous, and precipitation changes should not change suddenly between adjacent grids. Therefore, it is necessary to punish the sudden change of the spatial gradient in the prediction results. The expression is as follows:

[0121] ;

[0122] in, Y represents the precipitation image in the forecast results, and Respectively expressed in x and y The gradient operator in the spatial direction is used to limit the spatial smoothness of the predicted precipitation image and avoid local jumps.

[0123] Thus, it can be understood that this embodiment can integrate 3D convolution with the temporal attention mechanism, add a saliency guidance module and a physical consistency loss term, thereby significantly improving the recognition rate, prediction accuracy, and interpretability of extreme rainstorms.

[0124] It should be noted that this embodiment does not limit the specific loss threshold, and it is sufficient to ensure that the model is in a convergence state when the loss value is less than the loss threshold.

[0125] Step S30: determining the rainstorm risk weight at the spatial location corresponding to each spatial grid based on the historical rainstorm event dataset.

[0126] Specifically, in some embodiments, the historical rainstorm event dataset includes rainstorm events within a preset year. Figure 1 Step S30 includes but is not limited to the following steps:

[0127] Step S31, counting the frequency of occurrence of rainstorm events on each spatial grid and forming a frequency heat map;

[0128] Step S32: normalize the frequency heat map into a significance probability map, wherein the significance probability map is used to indicate the rainstorm risk weight of the spatial position corresponding to each spatial grid.

[0129] Specifically, the rainstorm events in a preset year correspond to hourly extreme event records for N years, and the specific value of N can be determined by a person skilled in the art based on actual conditions.

[0130] It should be noted that the frequencies of occurrence in this embodiment are calculated according to the following formula:

[0131] ;

[0132] in, is the grid point position of the spatial grid ( i , j ) number of heavy rains on is the total number of time steps (e.g., N years × 365 days × 24 hours). The frequency of occurrence of each grid point is counted to obtain a frequency heat map, which is then normalized into a significance probability map, and finally normalized into a significance probability map S.

[0133] Step S40: Fusing the rainstorm risk weight with the first feature tensor to obtain a second feature tensor with enhanced features.

[0134] Specifically, Figure 1 Step S40 includes but is not limited to the following steps:

[0135] Step S41, expanding the dimension of the saliency probability map based on the first feature tensor to obtain a reference saliency probability map, wherein the number of dimensions of the first feature tensor is the same as the number of dimensions of the reference saliency probability map;

[0136] Step S42: Perform pixel-by-pixel weighted fusion on the reference saliency probability map and the first feature tensor to obtain a second feature tensor.

[0137] In this embodiment, the rainstorm risk weights of the significance probability map S are fused with the first feature tensor to obtain a second feature tensor with enhanced features. The expression is as follows:

[0138] ;

[0139] ;

[0140] in, Represents a pixel-by-pixel channel penalty operation, which is equivalent to using historical experience probability to strengthen the model's attention mechanism and guide it to focus on high-risk areas in space.

[0141] Step S50 , performing a time series-based rolling decoding process on the second feature tensor to obtain a spatiotemporally continuous heavy rain prediction result.

[0142] Specifically, in some embodiments, Figure 1 Step S50 includes but is not limited to the following steps:

[0143] Step S51: Performing spatial dimension enhancement processing on the second feature tensor of each time step through the upsampling structure and convolution structure of the decoder to obtain a third feature tensor, wherein the spatial dimension of the third feature tensor is the same as the spatial dimension of the four-dimensional tensor;

[0144] Step S52, compressing the meteorological variable dimension of each third feature tensor to 1 through the second convolutional layer to obtain each fourth feature tensor;

[0145] In step S53, all fourth feature tensors are concatenated in chronological order to obtain a rainstorm prediction result.

[0146] Referring to the description of the above embodiment, the second feature tensor combines the spatiotemporal modeling features and the historical significance probability map information, and is the feature output with the most comprehensive expression capability in the entire backbone network. This embodiment uses the upsampling structure and convolution structure of the decoder to t The spatial dimension of the second feature tensor is improved to obtain the third feature tensor, where the spatial dimension of the third feature tensor is the same as that of the four-dimensional tensor, that is, the spatial dimension is restored and the spatial dimension is gradually improved to the original 1km resolution. The meteorological variable dimension of each third feature tensor is compressed to 1 (i.e., precipitation intensity) through the second convolution layer to obtain the meteorological variable dimension of each time step. t Finally, all the fourth feature tensors are spliced ​​in chronological order to obtain the rainstorm prediction result.

[0147] Specifically, the rainstorm prediction results of this embodiment include a high-resolution precipitation intensity map (unit: mm / h) and an hourly rolling prediction covering the next 1 to 6 hours.

[0148] Specifically, the grid dimension of this embodiment can be set to 1 km × 1 km through the decoder to meet the short-term forecast accuracy requirement.

[0149] It can be understood that the rainstorm prediction method based on four-dimensional spatiotemporal perception of this embodiment is applied to a rainstorm prediction system, which includes the following five functional modules: (1) a data encoding module, which is used to convert the original multi-source meteorological variables into a unified four-dimensional tensor; (2) a spatiotemporal modeling module: constructed by a target 3D convolutional model and an improved temporal cross-attention mechanism, which is used to capture the dynamic evolution characteristics of the four-dimensional tensor across time steps and height layers, that is, to obtain the first feature tensor; (3) a saliency guidance module, which is used to construct a saliency concept. The rate map is used to obtain the rainstorm risk weight of each spatial grid, which can guide the feature fusion operation to focus on potential extreme areas and enhance the model's focusing ability; (4) the physical consistency constraint module introduces physical constraint loss functions such as water vapor flux balance and energy conservation in the training target 3D convolution model stage to improve the physical interpretability of the model; (5) the prediction decoding module is used to fuse the output results of the spatiotemporal modeling module and the significance guidance module, and decode the fusion results to obtain the future multi-time precipitation intensity forecast map with a spatial resolution of up to 1 km, which can adapt to the time series requirements of minutes to hours.

[0150] In this way, this application enhances the model's ability to depict the spatial-temporal-hierarchical dynamic evolution of extreme rainstorms through four-dimensional spatiotemporal sequence modeling and significance guidance mechanisms. Combined with a variety of physical constraint loss functions, the model's prediction results have stronger physical consistency and credibility, adapting to multi-scale warning needs from minute to hour levels, and significantly improving the practicality and stability of precipitation forecasts. Compared with existing rainstorm prediction methods, this application pays more attention to the temporal continuity of precipitation evolution, and fully considers the changes in meteorological variables at different altitudes in the atmosphere and at different pressure layers, that is, realizing spatiotemporal perception of meteorological variables, which greatly improves the accuracy of rainstorm prediction results.

[0151] like Figure 2 As shown, Figure 2 : is a structural diagram of a control device provided in one embodiment of the present application. The present invention also provides a control device 200, comprising:

[0152] The processor 210 may be implemented as a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is configured to execute relevant programs to implement the technical solutions provided in the embodiments of the present application.

[0153] The memory 220 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 220 can store an operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 220 and is called by the processor 210 to execute the heavy rain prediction method based on four-dimensional space-time perception in the embodiments of this application.

[0154] Input / output interface 230, used to implement information input and output;

[0155] Communication interface 240, used to implement communication interaction between the apparatus and other devices, which can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WiFi, Bluetooth, etc.);

[0156] bus 250 , which transmits information between various components of the device (e.g., processor 210 , memory 220 , input / output interface 230 , and communication interface 240 );

[0157] The processor 210 , the memory 220 , the input / output interface 230 and the communication interface 240 are connected to each other in communication within the device via the bus 250 .

[0158] In addition, an embodiment of the present application further provides an electronic device, including the control device 200 of the above embodiment.

[0159] In addition, an embodiment of the present application also provides a storage medium, which is a computer-readable storage medium. The storage medium stores a computer program, and when the computer program is executed by a processor, it implements the above-mentioned heavy rain prediction method based on four-dimensional space-time perception.

[0160] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory optionally includes a memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of the above-mentioned networks include but are not limited to the Internet, an intranet, a local area network, a mobile communication network and a combination thereof. The device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and are located in one place, or may be distributed to multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment.

[0161] Those skilled in the art will appreciate that all or some of the steps and systems disclosed above can be implemented as software, firmware, hardware, or any suitable combination thereof. Some or all of the physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software may be distributed on computer-readable media, which may include computer storage media (or non-transitory media) and communication media (or transient media). As is well known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information, such as computer-readable instructions, data structures, program modules, or other data. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disks (DVDs) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by a computer. Furthermore, as is well known to those skilled in the art, communication media typically includes computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

[0162] The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the above implementation. Those skilled in the art can also make various equivalent modifications or substitutions under the shared conditions that do not violate the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of the present invention.

Claims

1. A rainstorm prediction method based on four-dimensional space-time perception, characterized in that: include: Acquire multiple initial meteorological variable parameters, and generate a four-dimensional tensor based on the initial meteorological variable parameters, wherein the four-dimensional tensor includes interrelated time series, vertical layers, spatial grids, and meteorological variables, and the vertical layers are the number of vertical atmospheric layers; Performing spatial feature extraction and temporal modeling on the four-dimensional tensor based on a trained target 3D convolution model to obtain a first spatiotemporal joint feature tensor, wherein the target 3D convolution model is trained based on the four-dimensional tensor and a joint constraint function, wherein the joint constraint function is composed of a water vapor conservation loss function, an energy conservation loss function, and a spatial gradient regularization constraint function; Determining a rainstorm risk weight at a spatial location corresponding to each of the spatial grids based on a historical rainstorm event dataset; Performing feature fusion on the rainstorm risk weight and the first feature tensor to obtain a second feature tensor with enhanced features; The second feature tensor is subjected to rolling decoding processing based on a time series to obtain a spatiotemporally continuous heavy rain prediction result.

2. The rainstorm prediction method based on four-dimensional space-time perception according to claim 1 is characterized in that: Generating a four-dimensional tensor based on the initial meteorological variable parameters, including: Resampling all of the initial meteorological variable parameters to a target spatial grid using a bilinear interpolation method to obtain a first intermediate variable; Unifying the original time resolutions of all the first intermediate variables to a target time frequency using a linear time interpolation method to obtain a second intermediate variable; Performing data correction and data cleaning on all the second intermediate variables to obtain third intermediate variables; Adjusting all the third intermediate variables into respective fourth intermediate variables, wherein the structured data format of the spatial grids corresponding to the respective fourth intermediate variables is the same; Based on each time step, concatenating the fourth intermediate variables at each vertical layer and each spatial grid to obtain a multi-channel tensor, wherein the time step belongs to the time series; The multi-channel tensors are stacked according to the time dimension to obtain the four-dimensional tensor.

3. The rainstorm prediction method based on four-dimensional space-time perception according to claim 1 is characterized in that: Based on the trained target 3D convolution model, spatial feature extraction and temporal modeling are performed on the four-dimensional tensor to obtain a first spatial-temporal joint feature tensor, including: For each meteorological variable in the four-dimensional tensor, based on each time step of the time series, the meteorological variable is mapped into a first intermediate tensor of a unified embedding dimension through a first convolutional layer or a linear layer; Input the intermediate tensor into the target 3D convolutional model, and generate a query Q vector, a key K vector, and a value V vector based on the first intermediate tensor at each time step and spatial position; Calculate the attention weight based on the Q vector and K vector at each time step; Performing weighted aggregation on the V vectors based on the respective attention weights to obtain respective second intermediate tensors; Stacking each of the second intermediate tensors according to the time dimension to obtain a third intermediate tensor; Perform residual processing on the third intermediate tensor to obtain the first feature tensor.

4. The rainstorm prediction method based on four-dimensional space-time perception according to claim 1 is characterized in that: The historical rainstorm event dataset includes rainstorm events in a preset year. Determining the rainstorm risk weight at the spatial location corresponding to each spatial grid based on the historical rainstorm event dataset includes: Counting the occurrence frequency of the rainstorm event on each spatial grid and forming a frequency heat map; The frequency heat map is normalized into a significance probability map, wherein the significance probability map is used to indicate the rainstorm risk weight of the spatial position corresponding to each of the spatial grids.

5. The rainstorm prediction method based on four-dimensional space-time perception according to claim 4 is characterized in that: Perform feature fusion on the rainstorm risk weight and the first feature tensor, The second feature tensor after feature enhancement is obtained, including: Expanding the dimension of the saliency probability map based on the first feature tensor to obtain a reference saliency probability map, wherein the number of dimensions of the first feature tensor is the same as the number of dimensions of the reference saliency probability map; The reference saliency probability map is weightedly fused with the first feature tensor pixel by pixel to obtain the second feature tensor.

6. The rainstorm prediction method based on four-dimensional space-time perception according to claim 1 is characterized in that: The second feature tensor is subjected to rolling decoding processing based on a time series to obtain a spatiotemporally continuous rainstorm prediction result, including: Performing spatial dimension enhancement processing on the second feature tensor at each time step through the upsampling structure and convolution structure of the decoder to obtain a third feature tensor, wherein the spatial dimension of the third feature tensor is the same as the spatial dimension of the four-dimensional tensor; Compressing the meteorological variable dimension of each of the third feature tensors to 1 through a second convolutional layer to obtain each fourth feature tensor; All of the fourth feature tensors are spliced ​​in chronological order to obtain the rainstorm prediction result.

7. The rainstorm prediction method based on four-dimensional space-time perception according to claim 1 is characterized in that: The target 3D convolutional model is trained according to the following steps: Inputting the four-dimensional tensor into an initial 3D convolution model and outputting an initial rainstorm prediction result; Calculating a loss value of the initial 3D convolutional model based on the joint constraint function and combining the initial rainstorm prediction result and the corresponding true prediction result; When the loss value is greater than a preset loss threshold, the model parameters of the initial 3D convolution model are iteratively adjusted based on the loss value and the loss threshold until the loss value is less than or equal to the loss threshold, thereby obtaining the target 3D convolution model.

8. A control device, characterized in that: It includes at least one control processor and a memory for communicating with the at least one control processor; the memory stores instructions that can be executed by the at least one control processor, and the instructions are executed by the at least one control processor to enable the at least one control processor to execute the heavy rain prediction method based on four-dimensional space-time perception as described in any one of claims 1 to 7.

9. An electronic device, characterized in that: Comprising the control device according to claim 8.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the heavy rain prediction method based on four-dimensional space-time perception as described in any one of claims 1 to 7.

Images