A precipitation inversion, short-impending precipitation prediction and rainstorm trajectory prediction method and system

Abstract

This invention discloses a method for precipitation inversion, short-term precipitation prediction, and heavy rain trajectory prediction, belonging to the field of precipitation prediction technology. The method constructs an input time series based on collected radar observation data. This input time series is then sequentially input into a three-dimensional spatiotemporal encoding-decoding model and a selective state-space model to obtain temporal features. The last sequence of the input time series is input into a physical evolution prior model to obtain a physical evolution prior field. Then, the terminal features of the three-dimensional spatiotemporal encoding-decoding model, the temporal features, and the physical evolution prior field are adaptively affinely fused and output to a multi-head output module to obtain the current precipitation inversion result, short-term precipitation prediction sequence, and heavy rain trajectory prediction sequence. The technical solution of this invention achieves significant improvements in precipitation inversion accuracy, short-term forecast stability, and heavy rain trajectory prediction capability.

Description

Technical Field

[0001] This invention relates to the field of precipitation prediction technology, and in particular to a method and system for precipitation inversion, short-term precipitation prediction, and rainstorm trajectory prediction. Background Technology

[0002] With the construction of multi-radar network business systems, in related rainfall forecasting technologies, due to the lack of spatial and physical consistency constraints, deep learning models are prone to noise "spicules" and boundary jumps in strong convective scenarios, resulting in unstable overall effects of precipitation inversion, short-term forecasting, and rainstorm trajectory tracking, making it difficult to reflect the physical laws of precipitation processes. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a method and system for precipitation inversion, short-term precipitation prediction, and rainstorm trajectory prediction, thereby improving the accuracy and stability of the overall performance of precipitation inversion, short-term prediction, and rainstorm trajectory tracking and prediction.

[0004] In a first aspect, the present invention provides a method for precipitation inversion, short-term precipitation prediction, and rainstorm trajectory prediction, particularly a method based on physical evolution priors and a selective state-space model. This method includes: collecting radar observation data; constructing an input time series based on the radar observation data; inputting the input time series into a three-dimensional spatiotemporal encoding-decoding model to obtain first feature information, and using the last slice of the first feature information as second feature information; inputting the last sequence of the input time series into a physical evolution prior model to obtain third feature information; inputting the first feature information into a selective state-space model to obtain fourth feature information; inputting the second, third, and fourth feature information into an affine fusion model to obtain fifth feature information; and inputting the fifth feature information into a multi-head output module to output the current precipitation inversion result, short-term precipitation prediction sequence, and rainstorm trajectory prediction sequence.

[0005] The technical solution of this invention introduces a priori physical evolution, which makes the movement path and morphological changes of precipitation fields more in line with physical laws. By using a selective state-space model, it enhances the ability to express the evolution of long-sequence echoes and improves the accuracy of identifying and predicting heavy precipitation events. As a result, it has achieved significant improvements in precipitation inversion accuracy, short-term forecast stability, and the ability to track and predict rainstorm trajectories.

[0006] In one possible implementation, the method further includes: preprocessing the input time series; wherein the input time series includes radar reflectivity Z and differential reflectivity Z0. DR Differential propagation phase rate K DPEcho top height (ETH), vertical liquid water content (VIL), correlation coefficient (CC); preprocessing of the input time series includes: Z, Z DR K DP z-score standardization is used; min-max normalization is used for ETH and VIL; CC is pruned and linearly scaled.

[0007] In one possible implementation, the second, third, and fourth feature information are input into an affine fusion model to obtain the fifth feature information, including: generating the scale and bias of the affine transformation based on the second feature information; performing an affine transformation on the fourth feature information, scale, and bias to obtain the sixth feature information; compressing the sixth feature information and performing residual fusion with the second feature information to obtain the seventh feature information; and expanding the third feature information channel and fusing it with the seventh feature information to obtain the fifth feature information.

[0008] In one possible implementation, the fifth feature information is input into a multi-head output module to output the current precipitation inversion result, short-term precipitation prediction sequence, and rainstorm trajectory prediction sequence. This includes: inputting the fifth feature information into a 1x1 convolution module to obtain the precipitation result at the current moment; and applying an activation function to the precipitation result to obtain the current precipitation inversion result.

[0009] In one possible implementation, the fifth feature information is input into the multi-head output module, which outputs the current precipitation inversion result, the short-term precipitation prediction sequence, and the rainstorm trajectory prediction sequence. This includes: using the current precipitation inversion result as the initial value, predicting the precipitation result at the next moment based on the fifth feature information and the precipitation result at the current moment, and obtaining a short-term precipitation prediction sequence for a preset time period.

[0010] In one possible implementation, the fifth feature information is input into a multi-head output module, which outputs the current precipitation inversion result, short-term precipitation prediction sequence, and rainstorm trajectory prediction sequence. This includes: inputting the fifth feature information into a 3x3 convolutional module and applying an activation function to obtain a central heatmap; inputting the fifth feature information into a 3x3 convolutional module to obtain rainstorm cell combination parameters; inputting the fifth feature information into a 3x3 convolutional module to obtain a planar displacement field; and inputting the central heatmap, rainstorm cell geometric parameters, and planar displacement field into a trajectory prediction model to obtain a rainstorm trajectory prediction sequence for a preset time period.

[0011] In one possible implementation, the central heat map, the geometric parameters of individual rainstorm cells, and the planar displacement field are input into the trajectory prediction model to obtain a rainstorm trajectory prediction sequence for a preset time period. This includes: constructing a rainstorm cell state vector sequence based on the central heat map, the geometric parameters of individual rainstorm cells, and the planar displacement field; inputting the rainstorm cell state vector sequence into the trajectory prediction model for time recursion and updating the hidden states; outputting the trajectory coordinates of each time step for the updated hidden states through a linear mapping layer to obtain the rainstorm cell trajectory sequence; and summing up the trajectory sequences of all rainstorm cells to obtain the rainstorm trajectory prediction sequence for the preset time period.

[0012] In one possible implementation, the method further includes: constructing a sample dataset, which includes the input time series, grid precipitation label sequences, rain gauge station-level label sequences, and rainstorm cell trajectory label sequences.

[0013] In one possible implementation, the method further includes: constructing a total loss function based on the current precipitation inversion loss, short-term precipitation prediction loss, station consistency loss, physical consistency loss, spatiotemporal consistency loss, and rainstorm trajectory prediction loss.

[0014] Secondly, embodiments of the present invention also provide a precipitation inversion, short-term precipitation prediction, and rainstorm trajectory prediction system based on a physical evolution prior and a selective state-space model. This system includes: a data acquisition module for acquiring radar observation data and constructing an input time series based on the radar observation data; a three-dimensional spatiotemporal encoding-decoding module for obtaining first feature information based on the input time series and using the last slice of the first feature information as second feature information; a physical evolution prior module for obtaining third feature information based on the last sequence of the input time series; a selective state-space module for obtaining fourth feature information based on the first feature information; an affine fusion module for obtaining fifth feature information based on the second, third, and fourth feature information; and a multi-head output module for obtaining the current precipitation inversion result, short-term precipitation prediction sequence, and rainstorm trajectory prediction sequence based on the fifth feature information.

[0015] Thirdly, the present invention also provides an electronic device, comprising: Memory stores computer-executable instructions non-transiently; The processor is configured to run computer-executable instructions. The computer-executable instructions are executed by the processor to implement the above-mentioned precipitation inversion, short-term precipitation prediction, and rainstorm trajectory prediction methods.

[0016] Fourthly, the present invention also provides a non-transient computer-readable storage medium, wherein the non-transient computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, implement the above-mentioned precipitation inversion, short-term precipitation prediction, and rainstorm trajectory prediction methods.

[0017] The technical solution of this invention has the following technical effects: physical evolution priors enable the movement path and morphological changes of precipitation fields to better conform to physical laws; the selective state-space model enhances the ability to express the evolution of long-sequence echoes, improving the accuracy of identifying and predicting heavy precipitation events. Thus, significant improvements are achieved in precipitation inversion accuracy, short-term forecast stability, and the ability to track and predict rainstorm trajectories. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 The flowchart illustrates a method for precipitation inversion, short-term precipitation prediction, and rainstorm trajectory prediction provided in this embodiment of the invention.

[0020] Figure 2 This is a schematic diagram of a precipitation inversion, short-term precipitation prediction, and rainstorm trajectory prediction system provided in an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of the structure of the 3D-UNet model provided in an embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram of the physical evolution prior model provided in an embodiment of the present invention.

[0023] Figure 5 This is a schematic diagram of the structure of the selective state-space model provided in an embodiment of the present invention.

[0024] Figure 6 This is a schematic diagram of the structure of the affine fusion model provided in an embodiment of the present invention.

[0025] Figure 7 This is a schematic diagram of the structure of the multi-head output model provided in an embodiment of the present invention.

[0026] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Figure 1 This invention provides a method for precipitation inversion, short-term precipitation prediction, and heavy rainfall trajectory prediction based on physical evolution priors and a selective state-space model, such as... Figure 1 As shown, it mainly includes the following steps.

[0029] S100: Collect radar observation data and construct an input time series based on the radar observation data.

[0030] The input time series includes radar reflectivity Z and differential reflectivity Z. DR Differential propagation phase rate K DP Echo top height (ETH), vertical liquid water content (VIL), and correlation coefficient (CC).

[0031] As one possible implementation, a preprocessing step for the input time series can also be included. Specifically, for Z, Z... DR K DP z-score standardization is used; min-max normalization is used for ETH and VIL; CC is pruned and linearly scaled.

[0032] One possible implementation method could also include quality control of the raw radar echo data. Specifically, this includes: identifying and removing ground clutter; suppressing biological echo interference; interpolating abnormal peak values ​​and missing values; and performing consistency correction on calibration errors among multiple radars. The network data is uniformly mapped onto an equidistant latitude and longitude grid with a spatial resolution of 0.5km × 0.5km, forming a grid domain of size H × W. Here, H × W represents the number of grid points in the east-west and north-south directions of the projected regular grid, i.e., the number of grid rows and columns.

[0033] For example, it includes the following steps.

[0034] S101. Construct the input time series X based on radar observation data.

[0035] At a certain prediction reference time The historical radar echo image sequence from the previous hour was selected as the input time series. , denoted as: (1) Each frame For size The six-channel two-dimensional image consists of: Channel 1: Radar reflectivity factor Z (dBZ); Channel 2: Differential reflectivity Z... DR (dB); Channel 3: Differential propagation phase rate K DP (° / km); Channel 4: Echo top height ETH (km); Channel 5: Vertical liquid water content VIL (kg·m⁻²); Channel 6: Correlation coefficient CC.

[0036] Input time series The overall dimensions are: ; The number 12 represents the number of time frames, with each frame spaced 5 minutes apart.

[0037] Step S102: Preprocessing of the input time series.

[0038] Different normalization methods are used for different physical quantities to make the numerical scales of each channel similar, which facilitates network training. For Z, Z DR K DP The channel is standardized using z-score: (2) in, , These are the mean and standard deviation of channel c, respectively; For ETH and VIL channels, min-max normalization is applied: (3) Fixed-range clipping and linear scaling are applied to the CC channel to ensure that the data falls within a stable range.

[0039] After normalization, it is still recorded as The overall size remains unchanged, for .

[0040] S200. Input the input time series into the 3D spatiotemporal encoding-decoding model 3D-UNet to obtain the first feature information, and use the last slice of the first feature information as the second feature information.

[0041] For example, such as Figure 3 As shown, it includes the following steps.

[0042] S201, Rearrangement of input time series.

[0043] The input time series X is rearranged into a form suitable for 3D convolution. Input size: 6 represents the number of channels, and 12 represents the time length; Rearranged input time series for: ; S202, Input Time Series The input is fed into the first convolutional layer to obtain feature p1; the first convolutional layer includes two 3D convolutional modules.

[0044] For example, input The convolutional layers are passed sequentially through two 3D Conv3D layers, where the first Conv3D layer has a kernel. Step size is Fill with Input channel 6, output channel 64; 2nd layer Conv3D: kernel Step size is Fill with The input channel is 64, the output channel is 64, and the output feature map is: .

[0045] S203. Input feature p1 into the first downsampling layer to obtain feature p2.

[0046] For example, feature p1 is spatially downsampled using the MaxPool3D max-pooling module to obtain feature p1. The kernel of MaxPool3D is... Step size is .

[0047] S204. Input feature p2 into the second convolutional layer to obtain feature p3, wherein the second convolutional layer includes a 3D convolutional module.

[0048] For example, features Features are obtained through a single 3D convolutional layer (Conv3D). .

[0049] S205. Input feature p3 into the second downsampling layer to obtain feature p4.

[0050] For example, features The input is downsampled into the MaxPool3D module to obtain features. .

[0051] S206. Input feature p4 into the bottleneck layer to obtain feature p5.

[0052] For example, for features Perform bottleneck layer convolution Conv3D to obtain features. .

[0053] S207. Input feature p5 into the first upsampling layer to obtain feature p6.

[0054] For example, for features Perform DeConv3D upsampling, set the output channels to 128, and restore the spatial resolution to [value missing]. , to obtain features .

[0055] S208. Input feature p3 and feature p6 into the first concatenation layer to obtain feature p7.

[0056] For example, features Features of jump connections After performing Concat concatenation, the features are obtained. .

[0057] S209. Input feature p7 into the third convolutional layer to obtain feature p8; wherein, the third convolutional layer includes a convolutional module.

[0058] For example, for features Perform Conv3D convolution to output features .

[0059] S210. Input feature p8 into the second upsampling layer to obtain feature p9.

[0060] For example, for features DeConv3D upsampling is performed to obtain features. .

[0061] S211. Input feature p9 and feature p1 into the second concatenation layer to obtain feature p10.

[0062] For example, features Features of jump connections Perform concatenation to obtain features. .

[0063] S212. Input feature p10 into the fourth convolutional layer to obtain feature p11; wherein, the fourth convolutional layer includes a convolutional module.

[0064] For example, for features Perform Conv3D convolution to obtain features. .

[0065] It is understandable that feature p11 is the first feature information.

[0066] S213. Select the last slice of feature p11 to obtain feature p12.

[0067] For example, from features Take the last frame in the time dimension (corresponding to) ), thus obtaining the two-dimensional features at the final time step. .

[0068] It is understandable that feature p12 is the second feature information.

[0069] S300. Input the last sequence of the input time series into the physical evolution prior model to obtain the third feature information.

[0070] For example, such as Figure 4 As shown, it may include the following multiple steps.

[0071] S301. Obtain the last sequence of the input time series.

[0072] For example, from the input time series Extract the last frame , as input to the prior model of physical evolution, where: .

[0073] S302, Displacement Feature Extraction and Displacement Field Prediction.

[0074] For example, will The input consists of two 3×3 convolutional layers with a stride of 1 and padding of 1, used to extract local features related to echo movement and directly output a two-dimensional displacement field. .

[0075] in, The two channels represent the displacement components in the horizontal and vertical directions, respectively, and are usually denoted as . ,Right now: (4)

[0076] S303. Input the last sequence of the input time series and the displacement field into the advection module to obtain the advection field.

[0077] For example, the last sequence With displacement field Input the GridSampleAdvect advect module together to obtain the advected radar field. .

[0078] Among them, GridSampleAdvect is a differentiable translation operator, which is based on the displacement field. right Perform coordinate offset and bilinear interpolation, with time step The advection operation is performed in seconds, satisfying:

[0079] in, For channel indexing, This represents the bilinear interpolation operator.

[0080] S304. Construct a residual subnet based on the last sequence of the input time series to obtain the residual. .

[0081] For example, to characterize the source-sink effects of precipitation formation and dissipation, from Start by constructing a residual subnet.

[0082] First, The input is another convolutional branch, first processed through a 3×3 convolution with a stride of 1 and padding of 1, resulting in... Then apply Activate, and you will get: ;Will Input a second 3×3 convolution, output the result of... Residual field of the same channel number .

[0083] in Used to characterize source and sink terms approximations caused by physical processes such as convection development, condensation and evaporation, outside of advection processes.

[0084] S305, Based on the advection field With residual Obtaining physical evolution priors .

[0085] For example, the advection field Yuanhui Residual Element-wise addition yields the output of the physical evolution prior module, EvoModule. .

[0086] Understandably, physical evolution is a priori. This is the third feature information.

[0087] S400. Input the first feature information into the selective state space model to obtain the fourth feature information.

[0088] For example, such as Figure 5 As shown, it may include the following multiple steps.

[0089] S401. Perform time averaging on feature p11 to obtain feature p13.

[0090] For example, features Along the time dimension Calculating the mean yields:

[0091] in Indicates the feature channel number (1-64), and the output size is: .

[0092] S402. Perform a convolution operation on feature p13 to obtain feature p14.

[0093] For example, After a Convolution with a stride of 1 maps the channel dimension from 64 to 128 to enhance feature representation and obtain the features. .

[0094] S403. Spatial flattening of feature p14 yields feature sequence p15.

[0095] For example, will In space dimension Flattened into a sequence of tokens, each token is a 128-dimensional vector, forming a sequence of length [length missing]. sequence: .

[0096] S404. Enhance the feature sequence p15 based on the selective state-space model to obtain the feature-enhanced sequence p16.

[0097] For example, will enter The selective state-space model MambaBlock (containing 2 layers of Mamba units with a hidden layer dimension of 128) yields feature-enhanced sequences that have stronger long-range relation representation capabilities in the spatiotemporal domain. .

[0098] S405. Restore the feature enhancement sequence p16 to obtain feature p17.

[0099] For example, the sequence Based on the original space dimensions Rearranged back into a two-dimensional feature map:

[0100] This is the final output of the selective state-space model (also known as the Mamba time series model).

[0101] It is understandable that feature p17 is the fourth feature information.

[0102] S500: Input the second, third, and fourth feature information into the affine fusion model to obtain the fifth feature information.

[0103] For example, such as Figure 6 As shown, it may include the following multiple steps.

[0104] S501. Generate the scale and bias of the affine transformation based on the second feature information.

[0105] For example, in order to implement Mamba features ( This step involves adaptive adjustment based on the characteristics of the main trunk. Passing through two respectively Convolution, generating the scale of affine transformations With bias : .

[0106] S502. Perform an affine transformation on the fourth feature information, scale, and bias to obtain the sixth feature information.

[0107] For example, for Performing an adaptive affine transformation yields:

[0108] in, This represents element-wise multiplication, and 1 guarantees... When the value is 0, the identity mapping is maintained, and the output size is: It is understandable that p18 is the sixth feature information.

[0109] S503. After compressing the sixth feature information, perform residual fusion with the second feature information to obtain the seventh feature information.

[0110] For example, After a Convolution compresses the number of channels from 128 to 64, resulting in: After channel compression With main characteristics Performing element-wise addition (residual fusion), we obtain: Among them, p20 is the seventh feature information.

[0111] S504. After expanding the third feature information channel, merge it with the seventh feature information to obtain the fifth feature information.

[0112] For example, physical evolution priors After a Convolution, expanding its channels to 64, yields: ; By adding element by element and Fusion to obtain the final multi-task shared representation: .in, This is the fifth feature information.

[0113] S600: Input the fifth feature information into the multi-head output module and output the current precipitation inversion result, short-term precipitation prediction sequence and rainstorm trajectory prediction sequence.

[0114] For example, such as Figure 7 As shown, it may include the following multiple steps.

[0115] S610. Input the fifth feature information into the 1x1 convolution module to obtain the precipitation result at the current time; apply an activation function to the precipitation result to obtain the current precipitation inversion result.

[0116] For example, the current precipitation inversion head (OHP head): will fuse features Enter one Convolution yields a linear output of the precipitation regression results at the current moment: To ensure the non-negativity of precipitation, for Apply Activate to obtain the current precipitation inversion results:

[0117] in, This represents the precipitation inversion field at the current moment, serving as the initial condition for future multi-step predictions.

[0118] S620. Using the current precipitation inversion result as the initial value, predict the precipitation result at the next moment based on the fifth feature information and the precipitation result at the current moment, and obtain a short-term precipitation prediction sequence for a preset time period.

[0119] An example, a future multi-step precipitation forecast head (QPF head): This head uses a recursive approach to predict the precipitation field for the next 0-3 hours (36 steps in 5-minute intervals). Let the initial state be: ; For the step( The following recursion is performed: The prediction field at the previous moment The recursively encoded features are obtained by mapping a 3×3 convolution to a 64-channel array. ; Will With fusion features By adding element-wise along the channel dimension, we obtain the basic features of the current step. ; right Apply Activation function to obtain nonlinear enhanced features ; Will The precipitation field is mapped to a single channel using a 1×1 convolution and then applied. Activate, obtain the first Step prediction results:

[0120] in ; The precipitation results obtained at 36 time steps The combination constitutes a precipitation forecast sequence for the next 0 to 3 hours.

[0121] Understandably, in the initial state, This refers to the precipitation result at the current moment (i.e., step k=0), which can be based on... Predict the precipitation results at the next moment (i.e., the k=1 step). Then, if the current time is step k=1, then inversion can be performed based on the current precipitation. Predict the precipitation outcome at the next moment (i.e., at step k=2). By analogy, the precipitation results at the next moment can be predicted based on the current precipitation results, thus obtaining a precipitation prediction sequence for 36 time steps from 0 to 3 hours in the future.

[0122] S630. Based on the fifth feature information, obtain the central heat map, the geometric parameters of the rainstorm cells, and the planar displacement field. Input the central heat map, the geometric parameters of the rainstorm cells, and the planar displacement field into the trajectory prediction model to obtain the rainstorm trajectory prediction sequence for a preset time period.

[0123] S631. Input the fifth feature information into the 3x3 convolutional module and apply the activation function to obtain the center heatmap.

[0124] For example, for features Using 3×3 convolution and Activate and output the center heatmap:

[0125] in, .

[0126] S632. Input the fifth feature information into the 3x3 convolution module to obtain the rainstorm single-unit combination parameters.

[0127] For example, using Convolution outputs the geometric parameters of a rainstorm cell:

[0128] in, The four channels are used to represent parameters such as the center offset, width, height, and orientation angle of the individual bounding box.

[0129] S633. Input the fifth feature information into the 3x3 convolution module to obtain the planar displacement field.

[0130] For example, using Convolution outputs a planar displacement field:

[0131] in, The two channels represent the displacement components in the horizontal and vertical directions, respectively, and are used to describe the motion trend of a single unit on a plane.

[0132] S634. Input the central heat map, the geometric parameters of the rainstorm cells, and the planar displacement field into the trajectory prediction model to obtain the rainstorm trajectory prediction sequence for a preset time period.

[0133] Specifically, after obtaining the central heat map Parameter diagram and displacement field Subsequently, the AssocGRU module of the gated cyclic unit can be used to perform time correlation and trajectory prediction on the rainstorm cells, and output the trajectory sequence for a future period of time.

[0134] For example, it may include the following steps.

[0135] S6341. Construct a sequence of state vectors at the single-unit level.

[0136] Specifically, at each time step, according to Candidate monomer centers are extracted from high-confidence locations (e.g., above a certain threshold), and the block diagram parameters are read at the corresponding locations. With displacement field The state vector of the individual entity at the current time is formed by concatenating the vectors:

[0137] in, Single-player index; : No. The center position of each individual unit on the grid; : No. Individual at time step The joint state vector.

[0138] S6342, Time recursion and association are performed based on the gated cyclic unit AssocGRU.

[0139] State sequence for each individual The hidden state update form is as follows: (Input is given by the AssocGRU module for time recursion.)

[0140] in, : No. Individual at time step The hidden state has a hidden layer dimension of 128; This indicates the update operation of the gated loop unit.

[0141] AssocGRU internally combines center confidence, box parameters, and displacement information to perform optimal matching between the predicted unit in the current frame and the ground truth / predicted unit in the previous frame, thereby achieving "identity preservation" and trajectory continuity of the unit.

[0142] S6343, Output the rainstorm trajectory prediction sequence TrajSeq.

[0143] Specifically, based on the updated hidden state The trajectory coordinates at each time step are output through a linear mapping layer. , forming a trajectory sequence:

[0144] The trajectories of all individual units are summarized as follows:

[0145] in, The time steps for trajectory prediction (e.g., the next 12 steps, 1 hour). : The number of rainstorm cells being tracked in the scene.

[0146] This step can be understood as the storm cell tracking head, which takes p* as input and includes three branches: central heat map, storm cell geometric parameters, and planar displacement field.

[0147] As one possible implementation, the method also includes: constructing a total loss function based on the current precipitation inversion loss, short-term precipitation prediction loss, station consistency loss, physical consistency loss, spatiotemporal consistency loss, and rainstorm trajectory prediction loss.

[0148] For example, it may include the following steps.

[0149] Step a: Basic regression loss of OHP and QPF.

[0150] Step a1: Huber+ intensity weighting.

[0151] For any pixel Define weighted loss:

[0152] in: True rainfall intensity (mm / h); The model predicts rainfall intensity, which can be based on the current moment. or future moments ; Weighting functions that segment by threshold or increase linearly, for example, for... Heavy rainfall can be weighted to mitigate data imbalance caused by extreme values.

[0153] The loss is:

[0154] in, : Transition (e.g.) ).

[0155] Step a2: Application in OHP and QPF.

[0156] Current OHP head loss (for) ):

[0157] Future 0–3h (36 steps) QPF head loss (for ):

[0158] Step b: Site consistency loss and ground correction.

[0159]

[0160] in: : No. The coordinates of each station; From the prediction of the grid field (which can be...) Or a certain In position Rainfall intensity obtained from bilinear sampling; Station rainfall intensity observation; Number of sites involved in the supervision.

[0161] Step c: Physical consistency loss.

[0162]

[0163] in: : The predicted first A frame of precipitation field can be understood as Time index format; Differentiable advection operator; The displacement / velocity field output by the displacement head (which can be obtained from the tracking head) (The sequence is given). Yuanhui Xiao The output is used to approximate microphysical sources and sinks and other systematic correction terms.

[0164] Step d: Spacetime consistent regularization.

[0165] Spatial smoothing constraints are applied to the current and future precipitation fields, in the form of total variational (TV) regularization:

[0166] in: : respectively represent along First-order difference operator for direction; :time In pixels The predicted rainfall intensity.

[0167] This loss can be applied to the current frame and all future frames, with a small weight, and is used to suppress glitches, ringing, and unreasonable sharp noise in the spatial neighborhood.

[0168] Step e: Rainstorm tracking branch loss.

[0169] This branch corresponds to the aforementioned Tracking head (S630) and includes four parts: center heatmap loss, box parameter loss, association matching loss, and trajectory.

[0170] Step e1: Focal Loss.

[0171] Central heat map Focal Loss is used:

[0172] in: Forecast Center Heatmap; : Whether the corresponding position is the truth label of the core center of the rainstorm; Focal Loss hyperparameter, can be selected .

[0173] Step e2: Box parameter loss (IoU Loss)

[0174] Parameter diagram of the box (or further decoded) Using IoU Loss, the typical form is:

[0175] in, : The parameters of the predicted bounding box output by the Tracking head; : The true bounding box parameters of a rainstorm cell.

[0176] IoU as defined by the standard:

[0177] in: :Predicted bounding box region ( (Decoded geometric region); : Real bounding box area ( (Decoded geometric region); Prediction box With real frame The area of ​​intersection; Prediction box With real frame The area of ​​the union of the sets.

[0178] Step e3: Association matching loss (Hungarian matching).

[0179] Hungarian matching establishes a one-to-one correspondence between predicted and actual cells. The matching cost typically consists of "center distance + IoU + velocity consistency".

[0180] in: : The distance between the prediction center and the truth center; : by displacement vector The difference between the derived velocity and the true velocity.

[0181] Step e4: Trajectory error.

[0182] Trajectory sequence output by AssocGRU L1 is used as the error metric:

[0183] in: : No. The rainstorm core will be in the future. Key points for step prediction (by) Extrapolation); : The corresponding truth trajectory point; The number of individual entities being tracked in the scene.

[0184] Step f: Total loss function.

[0185] By weighting and combining the above losses, we obtain the total training loss function of this invention:

[0186] in This is the loss weighting coefficient.

[0187] This invention provides a method for precipitation inversion, short-term precipitation prediction, and heavy rain trajectory prediction based on a physical evolution prior and a selective state-space model. The method includes: constructing an input time series based on collected radar observation data; sequentially inputting this input time series into a three-dimensional spatiotemporal encoding-decoding model and a selective state-space model to obtain temporal features; inputting the last sequence of the input time series into a physical evolution prior model to obtain a physical evolution prior field; and then adaptively affinely fusing the terminal features of the three-dimensional spatiotemporal encoding-decoding model, the temporal features, and the physical evolution prior field, and outputting the results to a multi-head output module to obtain the current precipitation inversion result, the short-term precipitation prediction sequence, and the heavy rain trajectory prediction sequence. This invention incorporates advection consistency constraints and source-sink corrections, making the movement path and morphological changes of the precipitation field more consistent with physical laws. Simultaneously, the Mamba state-space structure enhances the model's ability to express the evolution of long-sequence echoes, improving the identification and prediction accuracy of heavy precipitation events. Overall, this invention achieves significant improvements in precipitation inversion accuracy, short-term forecast stability, and heavy rain trajectory tracking and prediction capabilities.

[0188] Figure 2 This invention provides a precipitation inversion, short-term precipitation prediction, and rainstorm trajectory prediction system based on physical evolution priors and a selective state-space model. Figure 2 As shown, the system includes the following modules.

[0189] The data acquisition module is used to collect radar observation data and construct input time series based on the radar observation data; The three-dimensional spatiotemporal encoding-decoding module is used to obtain the first feature information based on the input time series and use the last slice of the first feature information as the second feature information. The physical evolution prior module is used to obtain the third feature information based on the last sequence of the input time series; A selective state space module is used to obtain fourth feature information based on the first feature information; The affine fusion module is used to obtain the fifth feature information based on the second, third, and fourth feature information. The multi-head output module is used to obtain the current precipitation inversion result, short-term precipitation prediction sequence and rainstorm trajectory prediction sequence based on the fifth feature information.

[0190] This invention also provides a training process for the aforementioned prediction method. Specifically, it may include the following steps.

[0191] Step a: Collect historical radar observation data and construct an input time series based on the historical radar observation data.

[0192] The specific implementation method is similar to the aforementioned step S100, and will not be repeated here.

[0193] Step b: Construct a sample dataset; this sample dataset includes the input time series, grid precipitation label sequence, rain gauge station-level label sequence, and rainstorm individual trajectory label sequence.

[0194] For example, it may include the following steps.

[0195] Step b1: Dataset partitioning.

[0196] Specifically, a sliding window method is used to construct samples in the time dimension. Each sample is a radar image within one hour, with a total of 12 frames. One sample corresponds to one time point T; each sample serves as an input. .

[0197] All samples were divided into training, validation and test sets in a ratio of 8:1:1 to ensure that different time periods and different weather processes were evenly distributed across the subsets.

[0198] Step b2: Create grid precipitation truth labels.

[0199] Specifically, a gridded precipitation true field is constructed using ground rain gauge observations as a reference:

[0200] The true value of precipitation at the current time T is denoted as: ;

[0201] The true precipitation sequence over 36 time steps from the next 0 to 3 hours is denoted as Each size is .

[0202] When using inverse distance weighting (IDW) to interpolate site data onto a grid, let the target grid point be... The surrounding rain gauges are gathered as The rainfall at the station was Then the grid rainfall estimate is:

[0203] Wherein the weights are:

[0204] The distance between the station and the target grid point. For exponential parameters, To avoid dividing by zero for small constants.

[0205] Step b3: Create rain gauge station-level labels.

[0206] Specifically, let the rain gauge set be... , No. Each station at time The observed value is The time-series precipitation observations for each station are recorded in the dataset for subsequent construction of the station consistency loss.

[0207] During training, the predicted precipitation field was analyzed. The location is then subjected to bilinear interpolation to obtain the predicted station value. ,and They jointly participated in the loss calculation.

[0208] Step b4: Create individual rainstorm trajectory tags.

[0209] Threshold and connected component analysis was performed on the true precipitation field at each moment to obtain the individual labels of severe convective rainstorms.

[0210] Step b41: For each frame Set one or more precipitation intensity thresholds (such as 16 mm / h, 32 mm / h), and extract grid points that meet the conditions to form candidate regions; Step b42: Perform connected component analysis on the candidate region to obtain multiple rainstorm cells. For each cell, calculate parameters such as center point coordinates, width, height, and major axis direction angle. Step b43: Project the center point onto On the grid, create a central heatmap label; use width, height, and angle as frame parameter labels; Step b44: Associate the rainstorm cells at consecutive moments to form a trajectory sequence, and record the location of the rainstorm center every 5 minutes in the next hour as the trajectory label of the rainstorm cell.

[0211] Each data sample in the sample dataset includes: input time series. Grid precipitation label sequence Rainfall station-level label sequence And the trajectory label sequence of individual rainstorm cells.

[0212] Step c: Train the prediction system model based on the sample dataset.

[0213] Similar to steps S200-S600 mentioned above, they will not be repeated here.

[0214] Understandably, the training set is used for updating model parameters, the validation set is used for model selection and hyperparameter tuning, and the test set is only used for independent performance evaluation after training is completed and does not participate in the model training process. During model training, the convergence of the validation set loss and related prediction evaluation metrics is used as the training termination condition. When the validation set performance no longer improves in several consecutive iterations, training stops and the optimal model parameters are saved.

[0215] The collaborative observation system of this invention is applied to electronic devices. Figure 3 A schematic diagram of the architecture of an electronic device suitable for implementing embodiments of the present invention is shown.

[0216] It should be noted that, Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0217] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by instructions (computer programs), or by instructions (computer programs) controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor. The electronic device of this embodiment includes a storage medium and a processor, wherein the storage medium stores multiple instructions that can be loaded by the processor to execute any step of the method provided in the embodiments of the present invention.

[0218] Specifically, the storage medium and the processor are electrically connected directly or indirectly to enable data transmission or interaction. For example, these components can be electrically connected to each other via one or more signal lines. The storage medium stores computer-executable instructions that implement data access control methods, including at least one software functional module that can be stored in the storage medium in the form of software or firmware. The processor executes various functional applications and data processing by running the software program and module stored in the storage medium. The storage medium can be, but is not limited to, Random Access Memory (RAM), Read-Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The storage medium stores the program, and the processor executes the program after receiving the execution instructions.

[0219] Furthermore, the software programs and modules within the aforementioned storage medium may also include an operating system, which may include various software components and / or drivers for managing system tasks (e.g., memory management, storage device control, power management, etc.) and can communicate with various hardware or software components to provide an operating environment for other software components. The processor may be an integrated circuit chip with signal processing capabilities. The aforementioned processor may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc., which can implement or execute the methods, steps, and logic flowcharts disclosed in this embodiment. The general-purpose processor may be a microprocessor or any conventional processor.

[0220] Since the instructions stored in the storage medium can execute the steps in any of the methods provided in the embodiments of the present invention, the beneficial effects of any of the methods provided in the embodiments of the present invention can be achieved, as detailed in the preceding embodiments, and will not be repeated here.

[0221] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for precipitation inversion, short-term precipitation prediction, and heavy rain trajectory prediction, characterized in that, The method includes: Collect radar observation data and construct an input time series based on the radar observation data; The input time series is input into a three-dimensional spatiotemporal encoding-decoding model to obtain the first feature information, and the last slice of the first feature information is used as the second feature information. The last sequence of the input time series is input into the physical evolution prior model to obtain the third feature information; The first feature information is input into the selective state space model to obtain the fourth feature information; The second feature information, the third feature information, and the fourth feature information are input into the affine fusion model to obtain the fifth feature information; The fifth feature information is input into the multi-head output module, which outputs the current precipitation inversion result, short-term precipitation prediction sequence, and rainstorm trajectory prediction sequence.

2. The method according to claim 1, characterized in that, The method further includes: preprocessing the input time series; The input time series includes radar reflectivity Z and differential reflectivity Z. DR Differential propagation phase rate K DP Echo top height (ETH), vertical liquid water content (VIL), correlation coefficient (CC); Preprocessing the input time series includes: For Z, Z DR K DP z-score standardization is adopted; Min-max normalization was applied to ETH and VIL; Crop and linearly scale the CC.

3. The method according to claim 1, characterized in that, The second feature information, the third feature information, and the fourth feature information are input into the affine fusion model to obtain the fifth feature information, including: The scale and bias of the affine transformation are generated based on the second feature information; The sixth feature information is obtained by performing an affine transformation on the fourth feature information, the scale, and the bias. The sixth feature information is compressed and then residually fused with the second feature information to obtain the seventh feature information; The third feature information channel is expanded and then fused with the seventh feature information to obtain the fifth feature information.

4. The method according to claim 1, characterized in that, The fifth feature information is input into the multi-head output module, which outputs the current precipitation inversion result, short-term precipitation prediction sequence, and rainstorm trajectory prediction sequence, including: The fifth feature information is input into a 1x1 convolution module to obtain the precipitation result at the current moment; An activation function is applied to the precipitation results to obtain the current precipitation inversion results.

5. The method according to claim 4, characterized in that, The fifth feature information is input into the multi-head output module, which outputs the current precipitation inversion result, short-term precipitation prediction sequence, and rainstorm trajectory prediction sequence, including: Using the current precipitation inversion result as the initial value, the precipitation result at the next moment is predicted based on the fifth feature information and the precipitation result at the current moment, thus obtaining a short-term precipitation prediction sequence for a preset time period.

6. The method according to claim 1, characterized in that, The fifth feature information is input into the multi-head output module, which outputs the current precipitation inversion result, short-term precipitation prediction sequence, and rainstorm trajectory prediction sequence, including: The fifth feature information is input into a 3x3 convolutional module and an activation function is applied to obtain a center heatmap; The fifth feature information is input into a 3x3 convolutional module to obtain the rainstorm individual combined parameters; The fifth feature information is input into a 3x3 convolution module to obtain a planar displacement field; The central heat map, the geometric parameters of the rainstorm cells, and the planar displacement field are input into the trajectory prediction model to obtain a rainstorm trajectory prediction sequence for a preset time period.

7. The method according to claim 6, characterized in that, The central heat map, the geometric parameters of the individual rainstorm cells, and the planar displacement field are input into the trajectory prediction model to obtain a rainstorm trajectory prediction sequence for a preset time period, including: A sequence of state vectors for rainstorm cells is constructed based on the central heat map, the geometric parameters of the rainstorm cells, and the planar displacement field. The sequence of state vectors of individual rainstorm cells is input into the trajectory prediction model for time recursion, and the hidden states are updated. The updated hidden state is processed by a linear mapping layer to output the trajectory coordinates of each time step, thus obtaining the trajectory sequence of the rainstorm unit. The trajectory sequences of all individual rainstorms are summarized to obtain the rainstorm trajectory prediction sequence for a preset time period.

8. The method according to claim 1, characterized in that, The method further includes: Construct a sample dataset, which includes the input time series, grid precipitation label sequence, rain gauge station-level label sequence, and rainstorm individual trajectory label sequence.

9. The method according to any one of claims 1-8, characterized in that, The method further includes: A total loss function is constructed based on the current precipitation inversion loss, short-term precipitation prediction loss, station consistency loss, physical consistency loss, spatiotemporal consistency loss, and rainstorm trajectory prediction loss.

10. A precipitation inversion, short-term precipitation prediction, and rainstorm trajectory prediction system, characterized in that, The system includes: The data acquisition module is used to collect radar observation data and construct an input time series based on the radar observation data. A three-dimensional spatiotemporal encoding-decoding module is used to obtain first feature information based on the input time series, and to use the last slice of the first feature information as second feature information. The physical evolution prior module is used to obtain third feature information based on the last sequence of the input time series; A selective state space module is used to obtain fourth feature information based on the first feature information; Affine fusion module is used to obtain fifth feature information based on the second feature information, the third feature information, and the fourth feature information; The multi-head output module is used to obtain the current precipitation inversion result, short-term precipitation prediction sequence and rainstorm trajectory prediction sequence based on the fifth feature information.

Images