Method, device and equipment for heavy rainfall nowcasting based on multi-task learning and motion task constraint

Abstract

The application discloses a heavy rainfall nowcasting method and device based on multi-task learning and motion task constraint, and equipment, first, the obtained original radar data is preprocessed; and based on the preprocessed radar data, a data set is constructed; then, a heavy rainfall nowcasting model of multi-task learning is established, including a multi-task prediction module and a motion task constraint module, wherein the multi-task prediction module is used for simultaneously predicting three radar variables and sharing network parameters, an adversarial generative network is introduced at the output end of the multi-task prediction module, and the motion task constraint module introduces a motion task constraint and is used for predicting a motion vector field containing translation and dissipation in historical radar observation; then, the heavy rainfall nowcasting model of multi-task learning is trained by using the constructed data set; finally, the heavy rainfall nowcasting model of multi-task learning is used for heavy rainfall nowcasting. The application can improve the accuracy of heavy rainfall nowcasting.

Description

Technical Field

[0001] This invention relates to the field of precipitation forecasting technology, and more specifically, to a method, apparatus, and equipment for nowcasting heavy rainfall based on multi-task learning and motion task constraints. Background Technology

[0002] Currently, nowcasting for severe convective weather primarily relies on radar observation data, employing storm identification and tracking technology and radar extrapolation methods to predict future precipitation. With the continuous improvement of data volume and computing power, and the rapid development of deep learning technology, new pathways have been opened for optimizing severe convective precipitation forecasts. The application of dual-polarization radar has also brought new ideas to forecasting work. This radar can transmit and receive horizontally and vertically polarized electromagnetic waves, acquiring microphysical parameters such as precipitation particle size, phase, and water content, thereby improving the accuracy of severe convective weather forecasts. By comprehensively utilizing deep learning technology and high-resolution radar data, future meteorological forecasting systems will be able to more accurately analyze and predict severe convective weather, effectively reducing the disaster risks brought by heavy rainfall.

[0003] In the prior art, document 1 with patent application number 202210632842.3 discloses a precipitation prediction method based on multi-encoder fusion of radar and precipitation information, including: using a radar information encoder in the precipitation prediction model to encode historical radar echo grid data to obtain radar-coded hidden state information; using a precipitation information encoder in the precipitation prediction model to encode historical precipitation grid data to obtain precipitation-coded hidden state information; concatenating the radar-coded hidden state information and the precipitation-coded hidden state information and performing layer normalization to obtain a new hidden state, which serves as the new initial state of the decoder in the precipitation prediction model. The initial input of the decoder is the output of the precipitation information encoder, and the decoder outputs the precipitation prediction result. The precipitation prediction model includes a radar information encoder, a precipitation information encoder, and a decoder. Reference 1's model relies solely on radar and precipitation observation data for fitting, with a loss function limited to graded weighted B-MSE. It lacks any physical constraints from fluid dynamics or atmospheric dynamics partial differential equations, learning only statistical correlations and failing to follow the physical laws governing actual precipitation evolution. Furthermore, it lacks prior meteorological physical constraints, relying on a purely data-driven black-box model, resulting in poor meteorological interpretability. Reference 2, patent application number 202311235963.5, discloses a precipitation forecasting method and system with a physically constrained prediction model. Its precipitation forecasting method constructs a prediction model and corresponding loss function, including MSE loss terms and convection-diffusion terms. However, it only uses the discretized convection-diffusion equation residuals as the loss regularization term, failing to incorporate other meteorological physical laws such as atmospheric mass conservation, water vapor budget, and convergence-divergence. Its physical constraints merely simplify the convection-diffusion equation, resulting in incomplete coverage of physical priors and a simplistic constraint method. Moreover, the loss function only includes MSE and the convection-diffusion term, neglecting classification losses and relying solely on regression losses for optimization, leading to insufficient forecast accuracy. Summary of the Invention

[0004] To solve, or at least partially solve, the above-mentioned technical problems, the present invention provides a method, apparatus and device based on multi-task learning and motion task constraint, which utilizes microphysical feature information in dual-polarization radar data to improve the accuracy of nowcasting of severe convective precipitation.

[0005] To achieve the above objectives, the first aspect of the present invention provides a method for nowcasting heavy rainfall based on multi-task learning and motion task constraints, comprising: Preprocess the acquired raw radar data; A dataset was constructed based on the preprocessed radar data; A multi-task learning model for heavy rainfall nowcasting is established, comprising a multi-task prediction module and a motion task constraint module. The multi-task prediction module simultaneously predicts three radar variables and shares network parameters. An adversarial generative network is introduced at the output of the multi-task prediction module to determine whether a single-frame forecast image is accurate and whether the echo changes over a continuous hour conform to the actual convection evolution. The motion task constraint module introduces motion task constraints to predict the motion vector field, which includes translation and dissipation components, in historical radar observations. The motion task constraint module includes a motion field encoder and a motion field decoder. The intermediate layer of the motion field encoder has a motion feature jump connection with the corresponding position of the multi-task learning module. The constructed dataset was used to train a multi-task learning model for nowcasting heavy rainfall. We use a well-trained multi-task learning model for nowcasting heavy rainfall to make nowcasts.

[0006] In one implementation, the acquired raw radar data is preprocessed, including: The acquired raw radar data is then visualized. The raw radar data is classified according to the visualization processing results and precipitation conditions, including wide-area and large-scale rainfall, local sharp and strong convection, uniform ordinary rainfall, and scattered special convection. Set a threshold to exclude samples from non-strong convection processes.

[0007] In one implementation, the preprocessed radar data includes four types of rainfall data, and a dataset is constructed based on the preprocessed radar data, including: A stratified sampling method was adopted, and each type of rainfall data was divided into training set, validation set and test set according to a certain proportion. The samples in each dataset were generated using a sliding window method.

[0008] In one implementation, the motion task constraint module specifically uses a preset optical flow algorithm to calculate the motion vector between adjacent frames. The calculation method is as follows:

[0009]

[0010] In the formula, Let the objective function be optical flow. This means finding the optimal pixel displacement that minimizes the target value. This represents the radar echo intensity at the position with coordinates (x, y) at time t; This represents the horizontal and vertical displacement of pixels in adjacent frames. Indicates the time interval between adjacent frames; Represents the spatial gradient of an image ， Represents the time gradient, Represents a time-series vector. A Represents the symmetric coefficient matrix of a quadratic polynomial; b,C These represent the coefficients of the linear term and the constant term of the polynomial, respectively.

[0011] In one implementation, the dataset includes a training set, which is used to train a multi-task learning heavy rainfall nowcasting model, including; The variable dimensions and contour dimensions contained in the training set samples are merged into a single channel to reconstruct the data and form the input sample. The input samples are fed into a multi-task learning model for nowcasting heavy rainfall, and the model is trained using a two-stage training method to obtain a trained model.

[0012] In one implementation, a two-stage training method is used to train the model, including: In Phase 1, the bi-discriminator network is frozen, and only the multi-task 3DST backbone is trained. The loss used includes multi-task mean square error loss and optical flow motion field supervision loss. The bi-discriminator network includes a spatial discriminator and a temporal discriminator. The multi-task 3DST backbone includes an encoder, a decoder, and a motion task constraint module. The motion task constraint module includes a motion field encoder and a motion field decoder. In Phase 2, the entire network is unfrozen, and the multi-task 3DST backbone and the bi-discriminator network are jointly trained. The losses used include multi-task mean square error loss, optical flow motion field supervision loss, generator loss, spatial discriminator loss, and temporal discriminator loss.

[0013] In one embodiment, the method further includes: Evaluate the effectiveness of the model.

[0014] Based on the same inventive concept, a second aspect of the present invention provides a heavy rainfall nowcasting device based on multi-task learning and motion task constraints, comprising: The data preprocessing module is used to preprocess the acquired raw radar data; The dataset building module is used to build datasets based on preprocessed radar data. The model building module is used to establish a multi-task learning model for heavy rainfall nowcasting. It includes a multi-task prediction module and a motion task constraint module. The multi-task prediction module is used to simultaneously predict three radar variables and share network parameters. An adversarial generative network is introduced at the output of the multi-task prediction module to determine whether the single-frame forecast map is true and whether the echo changes over a continuous hour conform to the actual convection evolution. The motion task constraint module introduces motion task constraints to predict the motion vector field, which includes translation and dissipation components, in historical radar observations. The motion task constraint module includes a motion field encoder and a motion field decoder. The intermediate layer of the motion field encoder has a motion feature jump connection with the corresponding position of the multi-task learning module. The model training module is used to train a multi-task learning model for nowcasting heavy rainfall using the constructed dataset. The heavy rainfall nowcasting module is used to perform nowcasting of heavy rainfall using a pre-trained multi-task learning heavy rainfall nowcasting model.

[0015] Based on the same inventive concept, a third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, it is used to implement the heavy rainfall nowcasting method based on multi-task learning and motion task constraints described in the first aspect.

[0016] Based on the same inventive concept, the fourth aspect of the present invention provides a heavy rainfall nowcasting device based on motion task constraints and generative adversarial networks, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it is used to implement the heavy rainfall nowcasting method based on multi-task learning and motion task constraints described in the first aspect.

[0017] Compared with the prior art, the advantages and beneficial technical effects of the present invention are as follows: This invention discloses a method for nowcasting heavy rainfall based on multi-task learning and motion task constraints. A multi-task learning model for nowcasting heavy rainfall is constructed, comprising a multi-task prediction module and a motion task constraint module. The multi-task prediction module simultaneously predicts three radar variables and shares network parameters. An adversarial generative network is introduced at the output of the multi-task prediction module to determine whether a single-frame forecast image is accurate and whether the echo changes over a continuous hour conform to the actual convection evolution. The motion task constraint module introduces motion task constraints to predict the motion vector field, which includes translation and dissipation components, from historical radar observations. The motion task constraint module includes a motion field encoder and a motion field decoder. The intermediate layer of the motion field encoder has a motion feature jump connection with the corresponding position of the multi-task learning module. This invention introduces motion field constraints at the input of the motion encoder, which are used to calculate the movement vectors and intensity dissipation changes between radars using a preset optical flow algorithm, generating a three-dimensional motion field feature map. Real convective motion features can be used as auxiliary features input to the network, and through motion field loss supervision, the model is forced to learn the advection, formation, and dissipation evolution patterns of rain clusters, constraining the predicted time-series motion trends to prevent distortion or misalignment; thereby improving the accuracy of local short-term convection and scattered thunderstorm location forecasts. Furthermore, a generative adversarial network is introduced at the output of the multi-task prediction module, which can reconstruct the fine echo structure of small- and medium-scale strong convection, further enhancing the model's performance. 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 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 This is a flowchart illustrating the heavy rainfall nowcasting method based on multi-task learning and motion task constraints provided in an embodiment of the present invention. Figure 2 This is a diagram illustrating the overall framework of the multi-task learning heavy rainfall nowcasting model in this embodiment of the invention. Figure 3 for Figure 2 The diagram shows the structure of the multi-task prediction module in the model shown. Figure 4 This is a schematic diagram of the mean horizontal reflectance factor visualization results of the precipitation process in an embodiment of the present invention. Figure 5 This is a schematic diagram showing the classification results and image features of four types of precipitation processes in an embodiment of the present invention; Figure 6This is a schematic diagram of the CSI index of different models predicting radar reflectivity for the next hour on a 1km contour surface in an embodiment of the present invention. Figure 7 This is a schematic diagram of the CSI index of different models predicting radar reflectivity for the next hour on a 3km contour surface in an embodiment of the present invention. Figure 8 This is a schematic diagram of the CSI index of different models predicting radar reflectivity for the next hour on a 7km contour surface in an embodiment of the present invention. Figure 9 This is a schematic diagram comparing the model of this invention with the 3D ST model. Detailed Implementation

[0020] Example 1 This embodiment provides a method for nowcasting heavy rainfall based on multi-task learning and motion task constraints. Please refer to [link to relevant documentation]. Figure 1 ,include: S1: Preprocess the acquired raw radar data.

[0021] Specifically, S1 is data preprocessing, which includes data visualization, classification, and data cleaning.

[0022] In one implementation, S1 can be achieved through the following steps: S1.1: Visualize the acquired raw radar data; S1.2: The raw radar data is classified according to the visualization processing results and precipitation conditions, including wide-area and large-scale rainfall, local sharp and strong convection, uniform ordinary rainfall, and scattered special convection; S1.3: Set a threshold to exclude samples from non-strong convection processes.

[0023] In the specific implementation process, in order to better understand and utilize the acquired radar data, visualization processing is first performed. All radar frames included in each complete rainfall process are averaged to generate a representative echo heat map, thereby intuitively distinguishing rainfall patterns.

[0024] This implementation uses three types of dual-polarization radar observation variables: horizontal reflectivity factor, differential reflectance, and differential phase shift. Since radar equipment clutter and ground object echoes generate numerous outliers exceeding the physically reasonable range, threshold truncation is first performed. The segmented truncation formula is as follows:

[0025] In the formula: , These are the maximum and minimum values ​​of the radar observations, respectively. These are truncated radar observations.

[0026] To adapt to the input of deep learning networks, the truncated observations are linearly mapped to the [0,1] interval, and the normalization formula is as follows:

[0027] In the formula: The normalized dimensionless values ​​are used as model inputs.

[0028] For each complete precipitation sequence, the altitude at 1km for all moments is... Perform time-series averaging on the field to generate a representative echo field:

[0029] in Let T be the radar echo of frame t, and T be the total number of frames for this precipitation segment. This represents the average echo field.

[0030] Based on the average echo pattern, precipitation processes are divided into four categories: widespread and large-scale rainfall, localized sharp and strong convection, uniform and ordinary rainfall, and scattered special convection.

[0031] Set a threshold to exclude samples from non-strong convection processes: Set The intensity threshold removes all rainfall segments with only weak stratospheric clouds and no strong echoes, thereby reducing the computational burden of invalid data and lowering model learning noise.

[0032] S2: Construct a dataset based on the preprocessed radar data.

[0033] Specifically, S2 involves the construction of the dataset. Based on the results of S1, the data is randomly divided into training, validation, and test sets within each type of precipitation process to construct a dataset for subsequent model training and evaluation.

[0034] In one implementation, S2 can be achieved through the following steps: A stratified sampling method was adopted, and each type of rainfall data was divided into training set, validation set and test set according to a certain proportion. The samples in each dataset were generated using a sliding window method.

[0035] In the specific implementation process, the first step is stratified sampling: the four types of rainfall are not mixed together and split. Each type is divided into training, parameter tuning and verification, and test sets according to a certain proportion to ensure that the proportion of wide-area, local and special types of rainstorms is balanced in the three sets of data, and to avoid the model being biased towards a certain type of rainfall.

[0036] Next is sample construction. This implementation uses a sliding window method to generate samples. Taking a single variable as an example, the sample generation process is as follows: (1) Define parameters: The number of frames in the i-th precipitation process is Sliding window length , Sliding step size S.

[0037] (2) Sample generation: For each precipitation event, starting from frame 1, a sliding window of length of [length missing] is used. The sliding window in The frame data moves with a sliding step size S. The first 10 frames of data for each sample are used as input and the last 10 frames of data are used as output. The sliding window moves throughout the precipitation process until it reaches the end of the precipitation process.

[0038] (3) Calculation of total sample size: First, calculate the number of samples that can be generated by a single rainfall process, and then add the samples generated by multiple rainfall processes to obtain the total number of samples in the entire dataset.

[0039] The sample size for a single precipitation event is calculated as follows:

[0040] The total number of samples in the entire dataset is calculated as follows:

[0041] in To generate sample numbers for a single precipitation event, L The total length of the sliding window (the sum of input and output frames); S Move the sliding window by the specified step size; N This represents the total number of samples after summarizing all precipitation events. Let be the total number of radar frames for the i-th precipitation process, and k be the total number of precipitation processes.

[0042] Total number of samples The functions of this invention include: (1) Balanced distribution of datasets Using the total sample size N as the base, the training, validation, and test samples are precisely divided according to a preset ratio (e.g., 0.65:0.15:0.2). At the same time, the proportion of the four types of rainfall samples is checked to avoid model bias caused by too few samples of a certain type of convection.

[0043] (2) Training hyperparameter settings The number of iterations per round is calculated using N, and the batch size, learning decay period, and early stopping waiting rounds are determined to adjust the training strategy to suit the dataset size.

[0044] (3) Normalization of loss and evaluation indicators RMSE, MAE, grid regularization loss, etc., all use the total number of samples as the denominator of the mean to ensure that different models and different experimental indicators can be compared horizontally.

[0045] (4) Unified standards for ablation tests All ablation experimental groups maintained a consistent total sample size N to eliminate interference from differences in sample size, and the changes in indicators only reflected the true gain of the network module.

[0046] (5) Experiment reproduction and computing power prediction With fixed window parameters and a total sample size N, the experimental results can be fully reproduced; at the same time, training memory and time consumption can be estimated based on N, supporting offline deployment of services.

[0047] S3: Establish a multi-task learning model for heavy rainfall nowcasting, including a multi-task prediction module and a motion task constraint module. The multi-task prediction module is used to simultaneously predict three radar variables and share network parameters. An adversarial generative network is introduced at the output of the multi-task prediction module to determine whether the single-frame forecast map is true and whether the echo changes over a continuous hour conform to the actual convection evolution. The motion task constraint module introduces motion task constraints to predict the motion vector field, which includes translation and dissipation components, in historical radar observations. The motion task constraint module includes a motion field encoder and a motion field decoder. The intermediate layer of the motion field encoder has a motion feature jump connection with the corresponding position of the multi-task learning module.

[0048] Specifically, S3 is the model construction, which builds a prediction framework based on multi-task learning and an encoder-decoder structure. On the basis of the above prediction framework based on multi-task learning, two network structures, motion field constraints and generative adversarial networks, are added to establish a heavy rainfall near-term prediction model based on multi-task learning and motion task constraints.

[0049] The model adopts a symmetrical encoding and decoding structure: the encoder progressively compresses the image size and extracts deep spatiotemporal features, while the decoder reverses the size to restore the original radar resolution. The features of each layer of the encoder are transmitted to the corresponding decoder layer through skip connections to avoid the loss of strong convection fine features during downsampling.

[0050] Multi-task synchronous prediction: The entire network shares a single encoder backbone, branching into three independent decoding branches, and performs predictions simultaneously. , and The three radar quantities allow the network to automatically learn the microscopic relationships between the three types of variables, and the total error is equal to the sum of the MSE errors of the three branches;

[0051] in Indicates the basic loss of multi-tasking; express Mean squared error loss for the predicted branch MSE; express Predict the branch MSE loss; express Predict the branch MSE loss.

[0052] Simultaneously, motion task constraints and a generative adversarial network (GAN) structure are introduced. Based on the aforementioned multi-task learning-based prediction framework, motion field constraints and a GAN are added to the network structure. The overall framework of the constructed prediction model is as follows: Figure 2 As shown.

[0053] Figure 2 The overall framework of the model includes a motion task constraint module and a multi-task prediction module. The motion task constraint module comprises a motion field encoder and a motion field decoder, while the multi-task prediction module includes one encoder and three prediction decoders, as detailed below. Figure 3 As shown.

[0054] Regarding sports field constraints An optical flow feature branch is added to the input of the 3DST encoder as a motion field constraint. Motion Skip-Connection is set in each layer of the encoder to synchronously send the solved translational or dissipated motion field features into the corresponding decoder layer in the multi-task prediction module. During training, motion field supervision loss is added to participate in the backpropagation of the total loss.

[0055] Meteorological physical equations are introduced to describe changes in radar echoes, while considering two major processes: horizontal movement and intensity dissipation.

[0056] in This represents taking the partial derivative with respect to time; This represents the radar echo translation vector field (representing the direction and speed of precipitation movement); Represents the spatial gradient operator; This represents the radar intensity dissipation field (representing changes in rainfall intensity).

[0057] Specifically, the Farneback dense optical flow algorithm is used to calculate the motion vector and intensity change between two adjacent radar frames on a grid-by-grid basis, generating a three-dimensional motion field feature map.

[0058]

[0059] In the formula Let the objective function be optical flow. This means finding the optimal pixel displacement that minimizes the target value. This represents the radar echo intensity at the position with coordinates (x, y) at time t; This represents the horizontal and vertical displacement of pixels in adjacent frames. Indicates the time interval between adjacent frames; Represents the spatial gradient of an image ， Represents the time gradient, Represents a two-dimensional joint vector. A Represents the symmetric coefficient matrix of a quadratic polynomial; b, C These represent the coefficients of the linear term and the constant term of the polynomial, respectively.

[0060] It should be noted that the coordinates (x, y) uniquely locate a specific grid point within the radar echo field. Used to represent the horizontal and vertical pixel displacements of the same grid point in the next frame, the two together constitute the rain cluster motion vector. It is a two-dimensional joint vector used when performing quadratic polynomial fitting. It combines the spatial horizontal coordinate x with the time coordinate t, but does not include y. It is only used for polynomial fitting of grayscale fields in local areas.

[0061] The 3D motion field features generated by the motion field constraints are fed into the encoder for feature extraction. The motion information is synchronously transmitted to each layer of the decoder, forcing the model to make predictions according to the real movement patterns of rainwater, thereby weakening the mean smoothing problem from the root.

[0062] Regarding the direct relationship between meteorological physics equations, the Farneback optical flow formula, and three-dimensional motion characteristic maps, the meteorological physics equations provide the theoretical basis. The formula for the meteorological mechanism of convective evolution, defined from a physical perspective, states that changes in radar echoes are caused by a translational vector field. Intensity dissipation field The shared determination is the theoretical basis for calculating the characteristics of the motion field. By optimizing the optical flow and solving for the optical displacement of each grid point pixel by pixel, the translation vector field in the formula is calculated. and intensity dissipation field The 3D motion field feature map is the calculation output carrier, which will contain the obtained solution. and The features are combined into a 3-channel 3D feature map (including temporal and spatial features), which is the motion field feature, and serve as the visualization and structured input features for the final input network.

[0063] Specifically, the role of motion field constraints includes: introducing atmospheric dynamics physical equations from radar echoes to add meteorological prior constraints to the pure data-driven network, compensating for the lack of physical laws in the black box of deep learning; using Farneback dense optical flow to calculate the movement vector and intensity dissipation changes of rain clusters in historical frames grid by grid, and inputting real convective motion features as auxiliary features into the network; through motion field loss supervision, forcing the model to learn the advection, formation and dissipation evolution of rain clusters, and constraining the predicted time series motion trends to avoid distortion and misalignment; and using motion feature skip connections to directly deliver fine-scale convective motion information to the decoder, compensating for the small and medium-scale convective movement details lost in downsampling. The achievable technical effects include: significantly reducing the inherent defects of model regression mean and suppressing overall blurring of forecast images from a physical mechanism perspective; improving the accuracy of local short-term convection and scattered thunderstorm location forecasts, reducing rain cluster shifts and morphological distortions; better continuity of long-term forecasts, avoiding unreasonable forecasts of rain clusters disappearing or suddenly increasing; and more accurate simulation of the movement of ice crystals and weak cloud systems at the 7km isohyet, reducing overall errors in high-altitude forecasts.

[0064] Unlike ordinary Skip-Connections, which only transmit raw radar spatiotemporal features, motion-type Skip-Connections are dedicated cross-layer connections for transmitting motion field features, forming a proprietary network structure for motion field constraints. Their data flow path is as follows: the motion field features calculated at each layer of the encoder bypass the conventional feature channel and are directly transmitted to the corresponding layer's decoder via a dedicated line. Ordinary Skip-Connections: only transmit raw radar image details (including...) , and The motion-based skip connections specifically transmit the translational and generation / dissipation motion vector information of rain clusters. The aim is to allow the decoder to simultaneously receive both spatial texture and temporal motion information when upsampling and reconstructing the echo, forcing the model to extrapolate rain cluster movement according to meteorological physics equations and avoiding forecast position deviations and temporal abrupt changes. Furthermore, the motion-based skip connections are embedded in a separate branch added in parallel between the 3DST encoder and decoder, existing alongside and not replacing the original image skip connections.

[0065] Regarding the structure of generative adversarial networks Mounted at the output of the multi-task decoder, it consists of two independent sub-networks: a spatial discriminator and a temporal discriminator, i.e., a dual-discriminator network. During the training phase, it alternates with the 3DST generator for updates, while during the inference phase, only the 3DST backbone is retained and the discriminator is discarded. It is equipped with a grid-regularized loss to synchronously constrain the strong echo region.

[0066] exist Two discriminant networks are added to the output of the predictive decoder: a spatial discriminator to determine whether a single-frame forecast image resembles the actual radar, and a temporal discriminator to determine whether the echo changes over a continuous hour conform to the actual convection evolution; simultaneously, a strong echo weighted grid regularization term is added to specifically constrain the heavy rainfall area; the objective of the two discriminators is to maximize the following:

[0067]

[0068] in, ; E[·] Represents the mathematical expectation; D(·) This represents the output of the spatial discriminator network; T(·) This represents the output of the time discriminator network; The parameter is The generator (3DST backbone); This represents historical radar observation samples; Represents the true value of the radar in the real future; Indicates the grid regularization weight coefficient; Indicates network regularization loss; H, W, N This represents the image height, width, and total number of samples; ⊙ This represents Hadamard element-wise multiplication; w(·) This represents a weighted mask for strong convection regions, used to amplify the weight of strong echo loss.

[0069] In practice, the basic judgment threshold is uniformly adopted as the strong convection judgment threshold. Using the true field of real radar A mask is generated based on the baseline, without the need for a prediction field, thus avoiding errors from interfering with weight allocation.

[0070] Formula for calculating the weight of each grid point For any spatial grid point (x, y), any temporal frame, and any contour surface, the weight values ​​are as follows:

[0071] The weighting is increased for regions with strong convection (in this implementation, it is set to 10, but the specific parameters can be adjusted according to the experiment). when If the area is identified as the core area of ​​a severe convective rainstorm, the weight is increased by 10 times; otherwise, for areas with stratus clouds, light rain, or no echo, the basic weight is set to 1, with normal constraints.

[0072] The complete generation steps are as follows: Input reference field: Take the future real-time data of each batch of training samples. Truth Tensor ; Grid-by-grid threshold judgment: Traverse all grid points of batch, time series, three-layer contour surface, and image width and height, and assign values ​​according to the above piecewise function; Generate a weighted mask tensor w of the same size: output a tensor with the same dimensions as the radar field, with 10 for strong convection grid points and 1 for the rest; Hadamard element-wise weighted loss: used in regularized loss The error is multiplied point by point by the mask; Synchronous application of mesh regularization loss : Amplify the loss contribution of strong echo regions, and prioritize correcting the rainstorm center error during training.

[0073] The generator loss, spatial discrimination loss, and temporal discrimination loss are designed separately. During training, the generator network and the two discriminant networks are updated alternately to continuously narrow the distribution gap between the forecast field and the real radar field and restore the fragmented strong convective echo. Mesh regularization loss The function is as follows:

[0074] The spatial discriminator loss function of a Generative Adversarial Network (GAN) is as follows:

[0075] in, This indicates the spatial discriminator loss. Indicates the network parameters of the spatial discriminator Indicates parameters Spatial discriminator; This represents the activation function.

[0076] The time-discriminatory loss function is as follows:

[0077] In the formula, Indicates the time discriminator loss. Indicates the network parameters of the time discriminator Indicates parameters Spatial discriminator.

[0078] In generative adversarial networks, the spatial discriminator's role is to predict single-frame contrast. Compared with real-time radar images, it determines whether the image texture and fragmented structure of strong convection are real; the time discriminator is responsible for the overall time series discrimination of 10 consecutive frames, constraining the one-hour echo evolution rhythm to conform to the real convection development logic; the grid regularization term weights the grid points of strong echoes above 35dBZ, prioritizing the optimization of forecast errors in the core area of ​​rainstorms; the adversarial loss is weighted and fused with multi-task MSE and motion field loss to bidirectionally optimize the clarity and temporal rationality of the generated images.

[0079] The achievable technical effects include: restoring the fine echo structure of strong convection at small and medium scales, solving the problems of strong center flattening and blurred outlines in traditional models; significantly improving the CSI critical success index; reducing false alarms and missed alarms, making the forecast of rainstorm area and intensity closer to the actual radar observation; layered adaptation of 1 / 3 / 7km contour features, with higher texture differentiation between thin clouds at high levels and heavy precipitation at low levels; and outputting forecast maps that conform to the visual interpretation habits of meteorological forecasters, directly supporting the use of short-term rainstorm and thunderstorm warnings.

[0080] S4: Use the constructed dataset to train the multi-task learning model for nowcasting heavy rainfall.

[0081] Specifically, S4 is the training of the model.

[0082] In one implementation, S4 can be achieved through the following steps: S4.1: Merge the variable dimensions and contour dimensions contained in the training set samples into a single channel to reconstruct the data and form the input sample; S4.2: Input the input samples into the multi-task learning heavy rainfall nowcasting model, and train the model using a two-stage training method to obtain the trained model.

[0083] Specifically, given that the model has multiple inputs containing multiple variables with multiple contour surfaces, in order to reduce the dimensionality of the model input, the variable dimension and the contour surface dimension are first merged into a single channel to reconstruct the data:

[0084] in The horizontal reflectance factor is the core predictive physical quantity of the radar. Differential reflectivity (dual-polarization radar observation); For differential phase shift (dual polarization radar observations); The vertical contour dimension of the radar surface; This refers to the moment of the last frame in the historical observation sequence; m For example, the number of time frames in a single segment. Corresponding to a radar observation / forecast duration of 60 minutes; Input the historical time series interval; For future forecast time series; For the 3DST spatiotemporal prediction model mapping function; This refers to all trainable weights and bias parameters of the model.

[0085] The input includes three height levels: 1km, 3km, and 7km. Each level is paired with three types of polarization data. The height and variable dimensions are merged and synthesized into nine input channels. These are then combined with the time, space, batch, height, and width dimensions to form the final five-dimensional input data, ultimately forming a five-dimensional array of [batch-time-channel-width-height]. The format is consistent with the video file, eliminating the need to calculate multiple sets of data separately, thus significantly improving computational speed. The model output only retains the height data from the three height levels. Reflectance data, corresponding to the predicted target.

[0086] S4.2 includes: In Phase 1, the bi-discriminator network is frozen, and only the multi-task 3DST backbone is trained. The loss used includes multi-task mean square error loss and optical flow motion field supervision loss. The bi-discriminator network includes a spatial discriminator and a temporal discriminator, and the multi-task 3DST backbone includes an encoder, a decoder, and a motion task constraint module. In Phase 2, the entire network is unfrozen, and the multi-task 3DST backbone and the bi-discriminator network are jointly trained. The losses used include multi-task mean square error loss, optical flow motion field supervision loss, generator loss, spatial discriminator loss, and temporal discriminator loss.

[0087] Specifically, in this invention, the generator of the model is a multi-task 3DST backbone (encoder, three decoders and motion constraint module), and the discriminator is a spatial discriminator and a temporal discriminator.

[0088] The core of the two-stage training strategy is to first train the backbone 3DST separately until convergence, and then release the discriminator for joint adversarial fine-tuning, thus avoiding the problems of backbone features easily collapsing and predictions being completely distorted during GAN synchronous training.

[0089] In the specific implementation process, Phase 1: Freeze the dual-discriminator network and pre-train only the MoG-3D backbone (120 epochs in total). (1) Parameter freeze operation Spatial discriminator Time discriminator All parameters are set to require_grad=False, gradients are not backpropagated and discriminator weights are not updated; only the 3DST encoder, three-way decoder, and optical flow motion field branch parameters are allowed to participate in the optimization.

[0090] (2) Losses incurred during this phase (no countermeasures)

[0091] For multi-tasking MSE loss, For the monitoring loss of optical flow sports field, These are the motion constraint weights.

[0092] (3) Training objectives Relying on radar physical priors and multi-polarization joint supervision, the backbone network learns the basic spatiotemporal laws of convection and advection, ensuring that the overall position and intensity of the output echo do not become disordered over a large area, thus achieving basic convergence; at the same time, it relies on motion-type jump connections to solidify the ability to extract rain cluster motion features.

[0093] (4) Convergence Criteria If the CSI and RMSE of the monitoring and validation set (30min / 60min) do not improve for 10 consecutive rounds, it is determined that the main trunk has converged and the second stage has begun.

[0094] Phase 2: Unfreeze the entire network and fine-tune the GAN joint adversarial approach (80 epochs in total) (1) Thawing operation Will With the parameter `requires_grad=True`, all parameters of the main branch, the motion field branch, and the dual discriminator can be updated, enabling complete joint training.

[0095] (2) Full total loss activated

[0096] This represents the overall joint total loss of the model, the final loss value during backpropagation during training. This represents the fundamental loss across multiple tasks, which is the sum of the mean square errors of the MSE of the three types of polarization radar quantities. The motion field monitoring loss is represented by the raindrop translation and dissipation field used to constrain optical flow calculations, which are consistent with the actual convection motion. The generator loss, which includes a strong convection-weighted grid regularization term, is used to optimize the texture of the predicted echo image. This represents the spatial discriminator loss, used to ensure consistency between single-frame forecast radar images and actual visual features. The time discriminator loss is used to constrain the evolution of the echo time series over a continuous hour to conform to the actual convection development pattern. This is the sports field loss weighting coefficient, used to adjust the proportion of physical motion constraints in the total loss. These are the generator's adversarial loss weights, used to control the degree of image detail optimization. The shared weight coefficients for the discriminator loss are used to simultaneously adjust the constraint strength of the spatial and temporal discriminant losses.

[0097] Simultaneously, generator adversarial loss, spatial discriminative loss, temporal discriminative loss, and strong convection weighted grid regularization term are introduced to fully constrain image texture and temporal continuity.

[0098] (3) Alternating update rule (standard game logic of GAN) Each batch of data undergoes two alternating backpropagation steps: ① Fix the generator backbone and only update the spatial and temporal discriminators to improve the ability to distinguish between true and false; ② Fix the discriminator, update the 3DST backbone generator, deceive the discriminator, and optimize the clarity of echo details. (4) Termination conditions If an early stopping strategy is set up, the validation set CSI will not increase for 12 consecutive rounds, and the optimal weight will be saved.

[0099] S5: Use a well-trained multi-task learning model for nowcasting heavy rainfall to perform nowcasting of heavy rainfall.

[0100] Specifically, S5 is the application of the model.

[0101] This invention proposes a nowcasting method for heavy rainfall based on multi-task learning and motion task constraints, which utilizes radar observation data from the previous hour (including horizontal reflectivity factor). Z H Differential reflectivity Z DR and phase shift K DP Using this as input, predict the next hour's... Z H Value. Specifically, to output the horizontal reflectance factor for the following hour. Z H The prediction results first involve a detailed classification of precipitation processes and the setting of appropriate thresholds to filter samples belonging to strong convective processes. Then, a semi-overlapping sliding window method is used to divide each precipitation process into multiple samples to better capture spatiotemporal variations. To handle the complex dependencies between radar observations, this invention employs a prediction model based on an encoder-decoder structure and multi-task learning, and adds motion field constraints and a generative adversarial network structure to this model. Furthermore, to expand the model's application scope, the Swin-Transformer's block embedding and window attention mechanism are extended to the three dimensions of spatiotemporal interaction, further improving the model's performance. Results show that the 3D prediction model constructed in this invention, based on multi-task learning and motion task constraints... D Swin-Transformer ( 3D ST The model demonstrated excellent performance on contour surfaces of 1km, 3km, and 7km.

[0102] In one implementation, the method further includes S6: evaluating the effectiveness of the model.

[0103] Specifically, the model trained in step S4 is used to predict the Z-value for the next hour under heavy rainfall conditions. H The CSI meteorological general evaluation index was selected, and a variety of classic spatiotemporal models were compared horizontally. Multiple ablation control experiments were carried out simultaneously to intuitively quantify the overall performance of the model and the individual gains of each module.

[0104] In the specific implementation process, the quantitative evaluation indicators mainly use the Critical Success Index (CSI, the core indicator for severe convection forecasting), the Root Mean Square Error (RMSE, the overall numerical error), and the Mean Absolute Error (MAE) to compare and evaluate the merits of each model.

[0105] The Critical Success Index (CSI) is calculated using the following formula:

[0106] In the formula Hit The number of grid points where strong echoes appear in both the actual situation and the forecast; Miss The number of grid points where strong echoes exist in the actual situation and where forecasts are missed; FalseAlarm This represents the number of grid points for real-time data without strong echoes and for forecasts without data.

[0107] Ablation control: Sequentially remove multi-polarization input, motion field constraints, and spatiotemporal constraints. GAN Three modules, each recording CSI The magnitude of the decrease provides a clear visual indication of the effectiveness of each module. Visual comparison: Outputs heatmaps for each forecast interval (6, 12...60 minutes), combining real-time radar data with 3D model data. DST predict, MoG- 3 DST The three images are displayed side by side, which makes it clear that the optimized model can retain more details of the rainstorm.

[0108] Layered error analysis: Statistical analysis of three contour levels at 1km, 3km, and 7km respectively. RMSE and CSI The indicators are used to quantitatively analyze the forecast deviations caused by ice crystals and thin clouds at an altitude of 7km, and to evaluate the forecast stability of the model at different altitude levels through vertical stratification.

[0109] The method of the present invention will be illustrated below through specific examples: Using a dual-polarization radar observation dataset, the ZH, ZDR, and KDP values ​​for the next hour (10 frames) are predicted based on radar observations (ZH, ZDR, KDP) from the past hour (10 frames). H Forecasts, along with radar observations including contour data at 1.0 km, 3.0 km, and 7.0 km, were used to build models and make precipitation predictions one hour later. Z H .

[0110] Step 1: Data Visualization and Classification. First, all frames of the 1km horizontal reflectance factor data for each precipitation event were averaged to obtain an averaged horizontal reflectance factor representing that precipitation event. After obtaining the averaged horizontal reflectance factor for the precipitation events, visualization processing was performed to present the averaged horizontal reflectance factors for different precipitation events (see details). Figure 4 ).

[0111] Based on the visualization results above, precipitation patterns are categorized into four types: spherical precipitation distribution (widespread precipitation), sharp radial precipitation (localized precipitation), a fuzzy classification between spherical and sharp radial precipitation distributions (common precipitation), and special categories of precipitation. The statistical data and image characteristics of the four main classification results are shown in Table 1 and [Table data missing]. Figure 5 .

[0112] Table 1. Statistics on the number of four types of precipitation events.

[0113] Step 2: Constructing the Prediction Dataset. Based on the precipitation process classification results above, construct a dataset for model training and evaluation. Within each precipitation process category, randomly partition the data into a training set, a validation set, and a test set, with a partition ratio of 0.65:0.15:0.2. This ratio is chosen to ensure sufficient data for model training, validation, and testing to evaluate model performance.

[0114] In both the training and validation sets, a semi-overlapping approach (i.e., S = 5) is used to alleviate the memory pressure of the sliding window process and significantly reduce the data storage space. In the test set, a sliding step size of 1 is used to simulate the real-time scrolling prediction requirements in practical applications.

[0115] Step 3: Build a heavy rainfall nowcasting model. Construct a prediction framework based on multi-task learning and an encoder-decoder structure; on the basis of the above prediction framework based on multi-task learning, add two network structures: motion field constraints and generative adversarial network, thereby establishing a 3D ST heavy rainfall nowcasting model based on multi-task learning and motion task constraints.

[0116] Step 4: Model Training. Given that the model has multiple inputs containing multiple variables with multiple contour surfaces, to reduce the model's input dimensionality, the variable dimension and contour surface dimension are first merged into a single channel to reconstruct the data, forming the input sample. This input sample is then fed into a multi-task learning model for heavy rainfall nowcasting, and a two-stage training method is used to train the model, resulting in a trained model.

[0117] Step 5: Model Prediction and Performance Evaluation. Using the model built in Step 3, predict the ZH value one hour after heavy rainfall. Select the CSI meteorological general evaluation index and compare multiple models to intuitively quantify the overall model performance and the individual gains of each module.

[0118] The results of this implementation method, comparing the improved model (ZhZdrKdp_MoG-3DST) that combines motion task constraints (Motion) and generative adversarial networks (GANs) with the 3D ST model (ZhZdrKdp_3DST), and comparing the multivariate model (ZhZdrKdp_MoG-3DST) with the univariate model (Zh_MoG-3DST), are as follows: Figures 6-9 As shown.

[0119] Figure 6 For the Z-shaped contour surface at 1km elevation H The predictions show that the MoG-3DST model exhibits a higher CSI index than the 3DST model at all times, indicating that the MoG-3DST model significantly outperforms the 3DST model in predicting ZH on a 1km contour surface. Similarly, the multivariate model (ZhZdrKdp) also significantly outperforms the univariate model (Zh) in terms of CSI index at all times, suggesting that the multivariate model utilizes dual-polarization radar observation data more effectively, resulting in better prediction performance. Figure 7 and Figure 8 These are predictions for contour surfaces at 3km and 7km respectively. For the Z-axis of the 3km and 7km contour surfaces... H The MoG-3DST model is predicted to show a higher CSI index than the 3DST model in most cases.

[0120] from Figure 9 As can be seen, the accuracy improved by 40.9% and 22.1% at the 30-minute and 60-minute prediction times, respectively, which further validates the advantages of the MoG-3DST model in ZH prediction. The ZhZdrKdp_MoG-3DST model exhibits the best performance in severe convective precipitation prediction, and future research can further optimize the model to adapt it to a wider range of meteorological applications.

[0121] Example 2 Based on the same inventive concept, this embodiment discloses a heavy rainfall nowcasting device based on multi-task learning and motion task constraints, comprising: The data preprocessing module is used to preprocess the acquired raw radar data; The dataset building module is used to build datasets based on preprocessed radar data. The model building module is used to establish a multi-task learning model for heavy rainfall nowcasting. It includes a multi-task prediction module and a motion task constraint module. The multi-task prediction module is used to simultaneously predict three radar variables and share network parameters. An adversarial generative network is introduced at the output of the multi-task prediction module to determine whether the single-frame forecast map is true and whether the echo changes over a continuous hour conform to the actual convection evolution. The motion task constraint module introduces motion task constraints to predict the motion vector field, which includes translation and dissipation components, in historical radar observations. The motion task constraint module includes a motion field encoder and a motion field decoder. The intermediate layer of the motion field encoder has a motion feature jump connection with the corresponding position of the multi-task learning module. The model training module is used to train a multi-task learning model for nowcasting heavy rainfall using the constructed dataset. The heavy rainfall nowcasting module is used to perform nowcasting of heavy rainfall using a pre-trained multi-task learning heavy rainfall nowcasting model.

[0122] Since the device in Embodiment 2 of this invention is the same device used in the heavy rainfall nowcasting method based on multi-task learning and motion task constraints in Embodiment 1, those skilled in the art can understand the specific structure and variations of this device based on the method described in Embodiment 1 of this invention, and therefore will not be repeated here. All devices used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0123] Example 3 Based on the same inventive concept, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a method for nowcasting heavy rainfall based on multi-task learning and motion task constraints, as described in Embodiment 1.

[0124] Since the computer-readable storage medium described in Embodiment 3 of this invention is the same computer-readable storage medium used in implementing the heavy rainfall nowcasting method based on multi-task learning and motion task constraints in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this computer-readable storage medium based on the method described in Embodiment 1 of this invention, and therefore will not be repeated here. All computer-readable storage media used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0125] Example 4 Based on the same inventive concept, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in Embodiment 1.

[0126] Since the computer device described in Embodiment 4 of this invention is the same computer device used to implement the heavy rainfall nowcasting method based on multi-task learning and motion task constraints in Embodiment 1 of this invention, those skilled in the art can understand the specific structure and variations of this computer device based on the method described in Embodiment 1 of this invention, and therefore will not be described again here. All computer devices used in the method of Embodiment 1 of this invention fall within the scope of protection of this invention.

[0127] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0128] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0129] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various modifications and variations to the embodiments of the invention without departing from the spirit and scope of the invention. Thus, if these modifications and variations of the embodiments of the invention fall within the scope of the claims of the invention and their equivalents, the invention also intends to include these modifications and variations.

Claims

1. A method for nowcasting heavy rainfall based on multi-task learning and motion task constraints, characterized in that, include: Preprocess the acquired raw radar data; A dataset was constructed based on the preprocessed radar data; A multi-task learning model for heavy rainfall nowcasting is established, comprising a multi-task prediction module and a motion task constraint module. The multi-task prediction module simultaneously predicts three radar variables and shares network parameters. An adversarial generative network is introduced at the output of the multi-task prediction module to determine whether a single-frame forecast image is accurate and whether the echo changes over a continuous hour conform to the actual convection evolution. The motion task constraint module introduces motion task constraints to predict the motion vector field, which includes translation and dissipation components, in historical radar observations. The motion task constraint module includes a motion field encoder and a motion field decoder. The intermediate layer of the motion field encoder has a motion feature jump connection with the corresponding position of the multi-task learning module. The constructed dataset was used to train a multi-task learning model for nowcasting heavy rainfall. We use a well-trained multi-task learning model for nowcasting heavy rainfall to make nowcasts.

2. The method for nowcasting heavy rainfall based on multi-task learning and motion task constraints as described in claim 1, characterized in that, The acquired raw radar data undergoes preprocessing, including: The acquired raw radar data is then visualized. The raw radar data is classified according to the visualization processing results and precipitation conditions, including wide-area and large-scale rainfall, local sharp and strong convection, uniform ordinary rainfall, and scattered special convection. Set a threshold to exclude samples from non-strong convection processes.

3. The method for nowcasting heavy rainfall based on multi-task learning and motion task constraints as described in claim 1, characterized in that, The preprocessed radar data includes four types of rainfall data. A dataset is constructed based on the preprocessed radar data, including: A stratified sampling method was adopted, and each type of rainfall data was divided into training set, validation set and test set according to a certain proportion. The samples in each dataset were generated using a sliding window method.

4. The method for nowcasting heavy rainfall based on multi-task learning and motion task constraints as described in claim 1, characterized in that, The motion task constraint module specifically uses a preset optical flow algorithm to calculate the motion vector between adjacent frames. The calculation method is as follows: In the formula, Let the objective function be optical flow. This means finding the optimal pixel displacement that minimizes the target value. This represents the radar echo intensity at the position with coordinates (x, y) at time t; This represents the horizontal and vertical displacement of pixels in adjacent frames. Indicates the time interval between adjacent frames; Represents the spatial gradient of an image ， Represents the time gradient, Represents a time-series vector. A Represents the symmetric coefficient matrix of a quadratic polynomial; b,C These represent the coefficients of the linear term and the constant term of the polynomial, respectively.

5. The method for nowcasting heavy rainfall based on multi-task learning and motion task constraints as described in claim 1, characterized in that, The dataset includes a training set, which is used to train a multi-task learning model for heavy rainfall nowcasting, including; The variable dimensions and contour dimensions contained in the training set samples are merged into a single channel to reconstruct the data and form the input sample. The input samples are fed into a multi-task learning model for nowcasting heavy rainfall, and the model is trained using a two-stage training method to obtain a trained model.

6. The method for nowcasting heavy rainfall based on multi-task learning and motion task constraints as described in claim 1, characterized in that, The model is trained using a two-stage training method, including: In Phase 1, the bi-discriminator network is frozen, and only the multi-task 3DST backbone is trained. The loss used includes multi-task mean square error loss and optical flow motion field supervision loss. The bi-discriminator network includes a spatial discriminator and a temporal discriminator. The multi-task 3DST backbone includes an encoder, a decoder, and a motion task constraint module. The motion task constraint module includes a motion field encoder and a motion field decoder. In Phase 2, the entire network is unfrozen, and the multi-task 3DST backbone and the bi-discriminator network are jointly trained. The losses used include multi-task mean square error loss, optical flow motion field supervision loss, generator loss, spatial discriminator loss, and temporal discriminator loss.

7. The method for nowcasting heavy rainfall based on multi-task learning and motion task constraints as described in claim 1, characterized in that, The method further includes: Evaluate the effectiveness of the model.

8. A heavy rainfall nowcasting device based on multi-task learning and motion task constraints, characterized in that, include: The data preprocessing module is used to preprocess the acquired raw radar data; The dataset building module is used to build datasets based on preprocessed radar data. The model building module is used to establish a multi-task learning model for heavy rainfall nowcasting. It includes a multi-task prediction module and a motion task constraint module. The multi-task prediction module is used to simultaneously predict three radar variables and share network parameters. An adversarial generative network is introduced at the output of the multi-task prediction module to determine whether the single-frame forecast map is true and whether the echo changes over a continuous hour conform to the actual convection evolution. The motion task constraint module introduces motion task constraints to predict the motion vector field, which includes translation and dissipation components, in historical radar observations. The motion task constraint module includes a motion field encoder and a motion field decoder. The intermediate layer of the motion field encoder has a motion feature jump connection with the corresponding position of the multi-task learning module. The model training module is used to train a multi-task learning model for nowcasting heavy rainfall using the constructed dataset. The heavy rainfall nowcasting module is used to perform nowcasting of heavy rainfall using a pre-trained multi-task learning heavy rainfall nowcasting model.

9. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the heavy rainfall nowcasting method based on multi-task learning and motion task constraints as described in any one of claims 1 to 7.

10. A heavy rainfall nowcasting device based on motion task constraints and generative adversarial networks, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the heavy rainfall nowcasting method based on multi-task learning and motion task constraints as described in any one of claims 1 to 7.

Images