Extreme short temporary rainfall forecasting method and system based on wavelet domain decoupling and multi-course learning

Abstract

The invention discloses an extreme short and temporary rainfall forecasting method and system based on wavelet domain decoupling and multi-course learning, and belongs to the technical field of meteorological big data processing and artificial intelligence deep learning. According to the invention, a meteorological image sequence is decomposed into a low-frequency approximate component and a high-frequency detail component through frequency domain decoupling, a double-branch deep neural network is adopted to extract features and predict corresponding frequency domain coefficients, and a rainfall prediction map is output through fusion of a spectral domain reconstruction and refining module. In addition, a multi-task course learning strategy is introduced during model optimization, and the loss weight is dynamically adjusted. The method is characterized in that modeling is moved to a wavelet domain to explicitly protect high-frequency details, local severe convection is accurately captured in combination with a double-branch architecture and a feature pyramid, the problem of traditional prediction fuzziness is effectively solved, the prediction definition and the extreme rainfall early-warning capacity are remarkably improved, high-fidelity and high-precision minute-level short-temporary rainfall prediction is achieved, and the prediction efficiency is improved. And reliable technical support is provided for meteorological disaster prevention and reduction.

Description

Technical Field

[0001] This invention relates to the fields of meteorological big data processing and artificial intelligence deep learning technology, and in particular to a method and system for extreme short-term precipitation forecasting using wavelet domain decoupling and multi-course learning. Background Technology

[0002] Precipitation forecasting is crucial for disaster preparedness, urban planning, and public safety; timely and accurate forecasts can effectively reduce loss of life and property. Traditional numerical weather prediction (NWP) relies on complex physical equations and numerical simulations, demonstrating high accuracy in medium- and long-term weather forecasts. However, due to its high computational complexity, NWP faces challenges in short-term nowcasting (typically 0-3 hours), including slow spin start-up and insufficient real-time performance. This is especially true in high-resolution output scenarios requiring minute-level updates, where computational resource demands increase significantly, making it difficult to meet the requirements of real-time disaster early warning.

[0003] In recent years, deep learning-based weather forecasting methods have gradually emerged, utilizing convolutional neural networks (CNNs) or recurrent neural networks (RNNs) to capture spatiotemporal features and generate prediction results within seconds. For example, ConvLSTM captures spatiotemporal correlations through gating mechanisms, while SimVP utilizes a pure CNN architecture to improve inference speed. However, existing deep learning short-term forecasting methods still face the following serious challenges: First, the pervasive smoothing effect, due to the long-tailed distribution of precipitation data, leads the optimization objective based on pixel-level mean square error to cause the model to tend to output the average value, resulting in blurred predicted images and loss of high-frequency texture details; second, insufficient ability to capture extreme events, traditional models struggle to learn rare extreme heavy precipitation (such as rainstorm cores and squall lines) features from a large number of light rain samples, leading to a severe underestimation of extreme values; and third, a lack of physical interpretability, as direct pixel-domain mixing modeling makes it difficult to distinguish between large-scale advection and local intensity generation and dissipation.

[0004] To address the aforementioned issues, there is an urgent need for a high-fidelity short-term precipitation prediction method that can explicitly decouple large-scale motion from small-scale texture and effectively overcome the smoothing effect. Summary of the Invention

[0005] The purpose of this invention is to address the problems existing in the prior art and to provide a method and system for forecasting extreme short-term precipitation using wavelet domain decoupling and multi-course learning. This invention achieves frequency domain decoupling by introducing discrete wavelet transform, utilizes a dual-branch deep neural network to model advection and intensity separately, and combines a multi-task course learning strategy to reduce prediction ambiguity, enhance extreme precipitation capture capabilities, improve structure fidelity, and meet the requirements for high-precision short-term forecasting.

[0006] To achieve the above-mentioned objectives, the present invention specifically adopts the following technical solution:

[0007] In a first aspect, the present invention provides a method for forecasting extreme short-term precipitation using wavelet domain decoupling and multi-course learning, comprising the following steps:

[0008] S1. Acquire historical meteorological images containing vertical integral liquid water variables and preprocess them. The preprocessed meteorological images form a meteorological image sequence containing spatiotemporal sequence information.

[0009] S2. The meteorological image sequence is input into the trained prediction model. After frequency domain decoupling of the meteorological image sequence, low-frequency approximation components and high-frequency detail components are obtained. Then, the two components are input into a two-branch deep neural network composed of an approximation network branch and a detail network branch for feature extraction and prediction. The approximation network branch takes the low-frequency approximation component as input and outputs the low-frequency prediction coefficients for future time steps. The detail network branch takes the high-frequency detail component as input and outputs the high-frequency prediction coefficients for future time steps. Finally, spectral domain reconstruction is performed based on the low-frequency prediction coefficients and the high-frequency prediction coefficients. The multi-path reconstruction map is then fused through a refinement module to output a precipitation prediction map. The prediction model adopts a multi-task course learning strategy for end-to-end optimization. By dynamically adjusting the loss weights of different frequency domain components, the prediction model is guided to gradually transition from learning the overall contour to refining local textures.

[0010] Based on the above scheme, each step can be implemented in the following preferred manner.

[0011] As a preferred embodiment of the first aspect mentioned above, the preprocessing process in step S1 includes downsampling the historical meteorological images in both time and space.

[0012] As a preferred embodiment of the first aspect, in step S2, the specific method for frequency domain decoupling of the meteorological image sequence is as follows: the meteorological image sequence is processed using two-dimensional discrete wavelet transform. For each preprocessed meteorological image frame, convolution and downsampling are performed using a low-pass filter and a high-pass filter to obtain a low-frequency approximate component and high-frequency detail components in the horizontal, vertical, and diagonal directions.

[0013] As a preferred embodiment of the first aspect above, in step S2, the detail network branch first maps the high-frequency detail components to the latent space by a multilayer perceptron to obtain latent features, and then extracts high-level semantic features from the latent features through a feature pyramid network. Then, the high-level semantic features are processed sequentially by two-dimensional convolution, normalization, and two-dimensional convolution to obtain high-frequency prediction coefficients.

[0014] The approximation network branch first encodes the low-frequency approximation components by the encoder to obtain the encoded original features. Then, the original features are processed by three-dimensional convolution to obtain the first feature. The first feature is concatenated with the original features and then processed by two-dimensional convolution to obtain the second feature. The second feature is then processed by multiple stacked hybrid modules to obtain the third feature. Finally, the third feature is decoded by the decoder to obtain the low-frequency prediction coefficients.

[0015] As a preferred embodiment of the first aspect mentioned above, each hybrid module is composed of a multilayer perceptron and a residual network cascaded in sequence.

[0016] As a preferred embodiment of the first aspect, in step S2, spectral domain reconstruction is performed using discrete wavelet inverse transform: the high-frequency prediction coefficients and low-frequency approximation components are reconstructed into a first reconstruction map, the low-frequency prediction coefficients and high-frequency detail components are reconstructed into a second reconstruction map, and the low-frequency prediction coefficients and high-frequency prediction coefficients are reconstructed into a third reconstruction map.

[0017] As a preferred embodiment of the first aspect mentioned above, in step S2, the refining module takes the first reconstructed image, the second reconstructed image, and the third reconstructed image as input. The first reconstructed image and the second reconstructed image are spliced ​​together to form the first fused reconstructed image. The first fused reconstructed image is processed by a residual network to obtain the fourth feature. The third reconstructed image is processed by a residual network to obtain the fifth feature. The fourth feature, the fifth feature, and the second reconstructed image are spliced ​​together to form the sixth feature. The sixth feature is processed by two-dimensional convolution to obtain the precipitation prediction map.

[0018] As a preferred embodiment of the first aspect above, in step S2, the loss function used by the prediction model during training is composed of four parts: the first part is the regularization loss. , is the first reconstructed graph Second Reconstruction Diagram Third Reconstruction Diagram The average value; the second part is the approximate branch loss. With dynamic weight scheduling function The product of the product of the predicted low-frequency coefficients and the actual low-frequency coefficients, the approximate branch loss is the zero-mean normalized cross-correlation loss of the low-frequency predicted coefficients and the actual low-frequency coefficients; the third part is the detail branch loss. The first part is the weighted mean square error of the high-frequency prediction coefficients and the true high-frequency coefficients; the fourth part is the pixel domain loss. , where represents the mean square error between the predicted precipitation map and the actual precipitation map; where the actual low-frequency coefficients and actual high-frequency coefficients are obtained by performing discrete wavelet transform on the actual precipitation map.

[0019] Furthermore, the above loss function The calculation method is shown in the following formula:

[0020]

[0021]

[0022] As a preferred embodiment of the first aspect mentioned above, the dynamic weight scheduling function takes the maximum value of two terms as the weight for approximating the branch loss; wherein, the first term is the difference between 1 and the first component, the first component is the ratio of the current iteration round to the preset total decay steps; and the second term is the preset minimum weight.

[0023] Furthermore, the dynamic weight scheduling function With the current iteration round The relevant calculation method is as follows:

[0024]

[0025] in, This indicates taking the maximum value; This represents the total number of decay steps. It has the minimum weight.

[0026] Secondly, this invention provides an extreme short-term precipitation forecasting system based on wavelet domain decoupling and multi-course learning, comprising:

[0027] The data processing module is used to acquire historical meteorological images containing vertical integral liquid water variables and perform preprocessing, which constitutes a meteorological image sequence containing spatiotemporal sequence information.

[0028] The result acquisition module is used to input meteorological image sequences into a trained prediction model. After frequency domain decoupling of the meteorological image sequences, low-frequency approximation components and high-frequency detail components are obtained. These two components are then input into a dual-branch deep neural network composed of an approximation network branch and a detail network branch for feature extraction and prediction. The approximation network branch takes the low-frequency approximation components as input to output low-frequency prediction coefficients for future time steps; the detail network branch takes the high-frequency detail components as input to output high-frequency prediction coefficients for future time steps. Finally, spectral domain reconstruction is performed based on the low-frequency and high-frequency prediction coefficients, and the multi-path reconstruction maps are fused through a refinement module to output a precipitation prediction map. The prediction model employs a multi-task course learning strategy for end-to-end optimization, guiding the prediction model from learning the overall contour to refining local textures by dynamically adjusting the loss weights of different frequency domain components.

[0029] Compared with the prior art, the present invention has the following advantages:

[0030] This invention boasts significant innovations and advantages. By shifting the modeling process from the pixel domain to the wavelet domain, it explicitly preserves high-frequency detail components, effectively solving the forecast ambiguity problem caused by "mean regression" in traditional deep learning models and significantly improving the clarity of forecast results. The dual-branch architecture combined with a feature pyramid design can accurately capture localized strong convective cells with randomness, significantly enhancing the early warning capability for extreme heavy precipitation (such as rainstorms and hail). The innovative multi-task course learning strategy effectively stabilizes the highly adversarial frequency domain learning process, preventing model training divergence. Through the above technical solutions, this invention achieves high-fidelity, high-precision minute-level short-term precipitation forecasts, providing strong technical support for real-time meteorological disaster prevention and mitigation. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the overall process of the method of the present invention;

[0032] Figure 2 This is a schematic diagram of the detailed network branch structure of the present invention;

[0033] Figure 3 This is a schematic diagram of the approximate network branch structure of the present invention;

[0034] Figure 4 This is a schematic diagram of the refining module structure of the present invention;

[0035] Figure 5 This is a line graph comparing the critical success index and extreme values ​​of the present invention under different prediction timeframes on the SEVIR dataset.

[0036] Figure 6 This is a line graph comparing the critical success index and high values ​​of the present invention under different prediction timeframes on the SEVIR dataset.

[0037] Figure 7 This is a line graph comparing the critical success index and average value of the present invention under different prediction timeframes on the SEVIR dataset.

[0038] Figure 8 This is a line graph showing the comparison of peak relative error under different prediction lead times on the SEVIR dataset for this invention;

[0039] Figure 9 This is a line graph showing the power comparison of the present invention under different prediction lead times on the SEVIR dataset;

[0040] Figure 10 This is a line graph showing the comparison of the root mean square error of the present invention under different prediction timeframes on the SEVIR dataset.

[0041] Figure 11 This is a graph showing the forecast results of the first and second comparative models under the heavy precipitation case event provided in this embodiment;

[0042] Figure 12 The diagram shows the third and fourth comparative models for heavy precipitation case events provided in this embodiment, and the forecast results of this invention.

[0043] Figure 13 The image shows the prediction results of the first and second comparative models under the squall line case event provided in this embodiment.

[0044] Figure 14 The diagram shows the third and fourth comparative models under the squall line case event provided in this embodiment and the prediction results of the present invention;

[0045] Figure 15 This is a graph showing the forecast results of the first and second comparative models for a tornado case event provided in this embodiment;

[0046] Figure 16 The diagram shows the third and fourth comparative models of the tornado case event provided in this embodiment and the forecast results of the present invention;

[0047] Figure 17 This is a system block diagram of the present invention. Detailed Implementation

[0048] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.

[0049] In the description of this invention, it should be understood that the terms "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include at least one of those features.

[0050] like Figure 1 As shown, in a preferred embodiment of the present invention, the extreme short-term precipitation forecasting method based on wavelet domain decoupling and multi-course learning includes the following steps S1-S2. The specific implementation process of each step will be described in detail below.

[0051] S1. Acquire historical meteorological images containing vertical integral liquid water variables and preprocess them. The preprocessed meteorological images constitute a meteorological image sequence containing spatiotemporal sequence information.

[0052] It should be noted that the preprocessing process in step S1 of the present invention includes downsampling the historical meteorological images in both time and space.

[0053] Furthermore, in step S1 of the present invention, the pixel values ​​in the historical meteorological images are mapped to the corresponding physical precipitation intensity values ​​through exponential transformation.

[0054] In this embodiment, the above-mentioned exponential transformation is specifically as follows:

[0055]

[0056] in, Represents pixel values ​​in historical weather images; This represents the mapped physical precipitation intensity value; It represents the natural logarithm.

[0057] In step S1 of this embodiment, the historical meteorological images used are derived from the SEVIR (Storm Event Imagery) dataset, which contains spatiotemporally aligned radar and satellite image sequences. This embodiment focuses on processing the vertically integrated liquid water (VIL) variable, a derivative of radar reflectivity, which effectively reflects the intensity and structure of storms. The historical meteorological images cover an area of ​​384×384 square kilometers with a temporal resolution of 5 minutes. In the preprocessing stage, to balance computational cost and information density, the historical meteorological images are downsampled spatially to a resolution of 128×128 and temporally to a 10-minute interval. The final input tensor, which contains meteorological images from the past 60 minutes (6 frames) (i.e., the aforementioned historical meteorological images), is used to predict precipitation conditions for the next 60 minutes (6 frames).

[0058] S2. Input the meteorological image sequence into the trained prediction model. After frequency domain decoupling of the meteorological image sequence, obtain the low-frequency approximate component and the high-frequency detail component, thereby explicitly separating the large-scale advection motion and local intensity fluctuations in the meteorological field. Then, input the two components into a two-branch deep neural network composed of an approximate network branch and a detail network branch for feature extraction and prediction. The approximate network branch takes the low-frequency approximate component as input and outputs the low-frequency prediction coefficients for future time steps. The detailed network branch takes high-frequency detail components as input and outputs high-frequency prediction coefficients for future time steps. Finally, spectral domain reconstruction is performed based on low-frequency and high-frequency prediction coefficients, and the multi-path reconstruction map is fused through a refinement module to eliminate artifacts and output a precipitation prediction map. The prediction model adopts a multi-task course learning strategy for end-to-end optimization. By dynamically adjusting the loss weights of different frequency domain components, the prediction model is guided to gradually transition from learning the overall outline to refining local textures.

[0059] It should be noted that, in step S2 of this invention, the specific method for frequency domain decoupling of the meteorological image sequence is as follows: the meteorological image sequence is processed using two-dimensional discrete wavelet transform (DWT). For each frame of the preprocessed meteorological image, convolution and downsampling are performed using a low-pass filter and a high-pass filter to obtain a low-frequency approximate component. High-frequency detail components in the horizontal, vertical, and diagonal directions .

[0060] In step S2 of this embodiment, each preprocessed meteorological image frame in the meteorological image sequence is decomposed using a biorthogonal wavelet basis. The decomposition process includes row filtering and column filtering, ultimately generating four components: a low-frequency approximation component, which contains the overall contour and smooth regions of the meteorological image, corresponding to large-scale advection in meteorology, and is used to characterize the macroscopic skeleton and advection trend of the storm system; and three high-frequency detail components, which respectively contain edge and texture information in the horizontal, vertical, and diagonal directions, corresponding to local intensity fluctuations and convective cell edges in meteorology, and are used to characterize local precipitation texture and extreme centers. Through this decomposition, the originally highly coupled precipitation field is separated into two physical components with distinct properties, providing a foundation for subsequent targeted modeling.

[0061] It should be noted that in step S2 of this invention, the parameters of the approximate network branch and the detail network branch are not shared. Specifically, as shown... Figure 2 As shown, the detail network branch first maps high-frequency detail components to the latent space using a multilayer perceptron to obtain latent features. Then, a feature pyramid network extracts high-level semantic features from the latent features. These high-level semantic features are then processed sequentially through two-dimensional convolution, normalization, and another two-dimensional convolution to obtain high-frequency prediction coefficients. Figure 3 As shown, the approximation network branch first encodes the low-frequency approximation components by the encoder to obtain the encoded original features. Then, the original features are processed by three-dimensional convolution to obtain the first feature. The first feature is concatenated with the original features and then processed by two-dimensional convolution to obtain the second feature. The second feature is then processed by multiple stacked hybrid modules to obtain the third feature. Finally, the decoder decodes the third feature to obtain the low-frequency prediction coefficients.

[0062] Furthermore, each hybrid module consists of a multilayer perceptron and a residual network cascaded together.

[0063] Furthermore, in this embodiment, both the encoder and decoder are two-dimensional convolutions. Other existing network structures can also be used, so there are no limitations in this invention.

[0064] In this embodiment, the detail network branch aims to capture the rapid generation and decay of high-frequency components. Due to the randomness and multi-scale characteristics of high-frequency information, this branch employs the existing multi-scale feature pyramid network (FPN) structure, resolving high-frequency textures at multiple resolution levels through bottom-up feature extraction and top-down semantic propagation. Specifically, the detail network branch first maps high-frequency detail components to the latent space, then extracts high-level semantic features through a bottom-up path, and propagates these high-level semantic features to low-level high-resolution feature maps through a top-down path. This cross-scale aggregation mechanism enables the model to simultaneously resolve coarse-grained structural details and fine-grained intensity edges, outputting future high-frequency prediction coefficients.

[0065] In this embodiment, the approximate network branch aims to capture the macroscopic evolution of low-frequency components. Since low-frequency information is spatially highly correlated and changes gradually, this branch employs a lightweight encoder-mixer-decoder structure: the encoder uses convolutional layers to extract features; the decoder corresponds to the encoder and also uses convolutional layers to extract features; the mixer includes 3D convolutions, residual networks, and multilayer perceptrons, enabling it to capture long-distance spatiotemporal dependencies within a large receptive field, simulating the overall movement trend of a storm system.

[0066] It should be noted that in step S2 of this invention, spectral domain reconstruction is performed using inverse discrete wavelet transform: the high-frequency prediction coefficients and low-frequency approximate components are reconstructed into a first reconstruction graph. The low-frequency prediction coefficients and high-frequency detail components are reconstructed into a second reconstruction image. Low-frequency prediction coefficients and high-frequency prediction coefficients Reconstructed into the third reconstruction diagram .

[0067] It should be noted that in step S2 of the present invention, as Figure 4 As shown, the refining module takes the first, second, and third reconstructed images as input. The first and second reconstructed images are concatenated to form the first fused reconstructed image. The first fused reconstructed image is processed by a residual network to obtain the fourth feature. The third reconstructed image is processed by a residual network to obtain the fifth feature. The fourth feature, the fifth feature, and the second reconstructed image are concatenated to form the sixth feature. The sixth feature is processed by two-dimensional convolution to obtain the precipitation prediction map. .

[0068] In this embodiment, to further improve image quality and eliminate block artifacts that may be generated during frequency domain synthesis, a refining module is designed. This refining module receives three inputs: independent reconstruction results of the approximate network branches (second reconstruction image). Independent reconstruction results of detail network branches (first reconstruction graph) ), and the direct inverse transform result of the combination of the two (third reconstruction graph) Internally, it contains two residual networks: one path learns the nonlinear interaction between low-frequency approximation components and high-frequency detail components, and the other path utilizes a third reconstructed graph as a strong structure prior to prevent prediction drift. Finally, a convolutional layer outputs a physically consistent future precipitation prediction map. .

[0069] It should be noted that in step S2 of this invention, the loss function used during the training of the prediction model is composed of four parts: the first part is the regularization loss. , is the first reconstructed graph Second Reconstruction Diagram Third Reconstruction Diagram The average value is used to constrain the three reconstructed graphs to tend towards consistency; the second part is the approximate branch loss. With dynamic weight scheduling function The product of the two factors, the approximate branch loss being the zero-mean normalized cross-correlation loss of the low-frequency predicted coefficients and the true low-frequency coefficients, is used to ensure the similarity of large-scale structures; the third part is the detail branch loss. The first part is the weighted mean square error of the high-frequency prediction coefficients and the true high-frequency coefficients, used to give different attention to details at different levels; the fourth part is the pixel domain loss. , where represents the mean square error between the predicted precipitation map and the actual precipitation map; where the actual low-frequency coefficients and actual high-frequency coefficients are obtained by performing discrete wavelet transform on the actual precipitation map.

[0070] Furthermore, the dynamic weight scheduling function The maximum value between the two terms is taken as the weight of the approximate branch loss; where the first term is the difference between 1 and the first component, and the first component is the ratio of the current iteration round to the preset total decay steps; the second term is the preset minimum weight.

[0071] In this embodiment, a loss function consisting of four parts is designed to stabilize the training process and enhance learning on extreme events. As shown in the following formula:

[0072]

[0073]

[0074] Dynamic weighted scheduling function With the current iteration round The relevant calculation method is as follows:

[0075]

[0076] in, This indicates taking the maximum value; This represents the total number of decay steps. It has the minimum weight.

[0077] In this embodiment, during the initial training phase, the dynamic weight scheduling function assigns higher weights to the approximate branch loss. This forces the model to first learn the shape distribution (advection) of the forecast storm, prioritizing the fitting of large-scale storm movement trajectories. As the number of training iterations increases, the dynamic weight scheduling function linearly decays to its minimum weight, gradually shifting the focus of model optimization to the detail branch loss and pixel domain loss, enabling the model to learn the internal texture and extreme intensity (intensity fluctuations) of the storm. This coarse-to-fine training approach simulates the human cognitive process, avoiding gradient divergence caused by directly optimizing complex objectives, and effectively improving the model's convergence stability under complex and extreme weather conditions.

[0078] Furthermore, the total decay steps and minimum weight can be set according to actual needs. The total decay steps are empirical values, rounded to the nearest integer, while the minimum weight must be a non-negative number. In this embodiment, ; .

[0079] To better demonstrate the specific implementation and technical effects of the present invention, the wavelet domain decoupling and multi-course learning extreme short-term precipitation forecasting method shown in steps S1 to S2 of the above preferred implementation is applied to a specific example.

[0080] Example

[0081] The specific implementation process of the wavelet domain decoupling and multi-course learning extreme short-term precipitation forecasting method used in this embodiment is as described above and will not be repeated here. For ease of description, the above extreme short-term precipitation forecasting method is simply referred to as the method of this invention, and the prediction model used is denoted as WADE-Pre.

[0082] Regarding the implementation details and advantages of this embodiment, the design of this invention emphasizes physical interpretability and high fidelity. Unlike traditional models that compress all information into a black box, the method of this invention uses wavelet decoupling, and the model structure can be consistent with the "advection + development" theory in meteorology.

[0083] Experiments show that, on the SEVIR dataset, the method of this invention significantly outperforms state-of-the-art precipitation forecasting models such as ConvLSTM, SimVP, and AlphaPre in terms of forecasting scores for extreme precipitation. Figure 5 , Figure 6 , Figure 7 , Figure 8 , Figure 9 , Figure 10As shown. Especially in the later stages of prediction (+60 minutes), the images generated by the method of this invention still maintain sharp edges and clear cores of heavy precipitation, without the "cloud dissipation" or "blurring" phenomena commonly seen in traditional models, demonstrating its significant advantage in solving the smoothing effect.

[0084] To verify the effectiveness of the method of this invention, this embodiment uses the SEVIR validation set for testing, with a prediction lead time of 60 minutes into the future. The prediction results are shown in Table 1. In Table 1, ↑ represents a higher value and a lower value and a better value, respectively. U-Net, ConvLSTM, SimVP, Earthfarseer, and AlphaPre are compared. CSI-M, CSI-E, CSI-H, RMSE, SSIM, and PSD are used as evaluation indicators. The best value for each evaluation indicator is bolded, and the second best value is underlined. Specifically, CSI-M represents the average CSI, and CSI-H represents the CSI for heavy precipitation (…). CSI-E indicates CSI for extreme precipitation ( SSIM represents the similarity of forecast results, RMSE represents the root mean square error of numerical accuracy, and PSD represents the power spectral density error, which measures image sharpness.

[0085] Table 1. Average index values ​​of different models at 6 time periods

[0086] As shown in Table 1, compared to the baseline model, the method of this invention achieves excellent results on all indices. Particularly for the CSI-E index of extreme precipitation, the method of this invention improves upon the suboptimal AlphaPre model by approximately 46%, fully demonstrating the powerful ability of the bi-branch decoupling design to capture rare extreme events. Furthermore, the lowest PSD error indicates that the images generated by the method of this invention have the best spectral fidelity, i.e., the clearest and most realistic texture closest to true radar echoes.

[0087] exist Figure 11 , Figure 12 , Figure 13 , Figure 14 , Figure 15 , Figure 16 This implementation example demonstrates a visual forecast comparison of three meteorological events: squall line, heavy precipitation, and tornado. Experimental results show that this invention is highly competitive in nowcasting of extreme precipitation, particularly in the accurate prediction of extreme values ​​over long time periods, and even more accurate in predicting the morphology of precipitation distribution. Through wavelet domain decoupling and multi-task learning, the smoothing effect in deep learning forecasting is effectively addressed, providing clearer and more accurate warnings of severe weather.

[0088] It should also be noted that the extreme short-term precipitation forecasting method based on wavelet domain decoupling and multi-course learning in the above embodiments can essentially be executed by a computer program or module. Therefore, similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides an extreme short-term precipitation forecasting system based on wavelet domain decoupling and multi-course learning, corresponding to the extreme short-term precipitation forecasting method based on wavelet domain decoupling and multi-course learning provided in the above embodiments, such as... Figure 17 As shown, it includes:

[0089] The data processing module is used to acquire historical meteorological images containing vertical integral liquid water variables and perform preprocessing, which constitutes a meteorological image sequence containing spatiotemporal sequence information.

[0090] The result acquisition module is used to input meteorological image sequences into a trained prediction model. After frequency domain decoupling of the meteorological image sequences, low-frequency approximation components and high-frequency detail components are obtained. These two components are then input into a dual-branch deep neural network composed of an approximation network branch and a detail network branch for feature extraction and prediction. The approximation network branch takes the low-frequency approximation components as input to output low-frequency prediction coefficients for future time steps; the detail network branch takes the high-frequency detail components as input to output high-frequency prediction coefficients for future time steps. Finally, spectral domain reconstruction is performed based on the low-frequency and high-frequency prediction coefficients, and the multi-path reconstruction maps are fused through a refinement module to output a precipitation prediction map. The prediction model employs a multi-task course learning strategy for end-to-end optimization, guiding the prediction model from learning the overall contour to refining local textures by dynamically adjusting the loss weights of different frequency domain components.

[0091] It is understood that the wavelet domain decoupling and multi-course learning extreme short-term precipitation forecasting method described in S1~S2 above can essentially be implemented by a computer program. Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer program product corresponding to the wavelet domain decoupling and multi-course learning extreme short-term precipitation forecasting method provided in the above embodiments. This product includes a computer program / instruction that, when executed by a processor, can implement the wavelet domain decoupling and multi-course learning extreme short-term precipitation forecasting method as described in the above embodiments.

[0092] Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer electronic device corresponding to the wavelet domain decoupling and multi-course learning extreme short-term precipitation forecasting method provided in the above embodiments, which includes a memory and a processor;

[0093] The memory is used to store computer programs;

[0094] The processor is configured to implement the extreme short-term precipitation forecasting method based on wavelet domain decoupling and multi-course learning in the above embodiments when executing the computer program.

[0095] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0096] Therefore, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer-readable storage medium corresponding to the wavelet domain decoupling and multi-course learning extreme short-term precipitation forecasting method provided in the above embodiments. The storage medium stores a computer program, which, when executed by a processor, can realize the wavelet domain decoupling and multi-course learning extreme short-term precipitation forecasting method in the above embodiments.

[0097] It is understood that the aforementioned storage media may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Furthermore, the storage media may also be various media capable of storing program code, such as USB flash drives, external hard drives, magnetic disks, or optical discs.

[0098] It is understood that the processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0099] It should also be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. In the embodiments provided in this application, the division of steps or modules in the system and method is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules or steps may be combined or integrated together, and a module or step may also be split.

[0100] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. An extreme short-term precipitation forecasting method of wavelet domain decoupling and multi-course learning, characterized in that, Includes the following steps: S1. Acquire historical meteorological images containing vertical integral liquid water variables and preprocess them. The preprocessed meteorological images form a meteorological image sequence containing spatiotemporal sequence information. S2. The meteorological image sequence is input into the trained prediction model. After frequency domain decoupling of the meteorological image sequence, low-frequency approximation components and high-frequency detail components are obtained. Then, the two components are input into a two-branch deep neural network composed of an approximation network branch and a detail network branch for feature extraction and prediction. The approximation network branch takes the low-frequency approximation component as input and outputs the low-frequency prediction coefficients for future time steps. The detail network branch takes the high-frequency detail component as input and outputs the high-frequency prediction coefficients for future time steps. Finally, spectral domain reconstruction is performed based on the low-frequency prediction coefficients and the high-frequency prediction coefficients. The multi-path reconstruction map is then fused through a refinement module to output a precipitation prediction map. The prediction model adopts a multi-task course learning strategy for end-to-end optimization. By dynamically adjusting the loss weights of different frequency domain components, the prediction model is guided to gradually transition from learning the overall contour to refining local textures.

2. The method of claim 1, wherein the wavelet domain decoupling and multi-course learning for extreme short-term precipitation forecast is characterized by, The preprocessing process in step S1 includes downsampling historical meteorological images in both time and space.

3. The method of claim 1, wherein the wavelet domain decoupling and multi-course learning for extreme short-term precipitation forecast is characterized by, In step S2, the specific method for frequency domain decoupling of the meteorological image sequence is as follows: the meteorological image sequence is processed by two-dimensional discrete wavelet transform. For each preprocessed meteorological image frame, convolution and downsampling are performed using low-pass and high-pass filters to obtain a low-frequency approximate component and high-frequency detail components in the horizontal, vertical, and diagonal directions.

4. The extreme short-term precipitation forecasting method based on wavelet domain decoupling and multi-course learning as described in claim 1, characterized in that, In step S2, the detail network branch first maps high-frequency detail components to the latent space by a multilayer perceptron to obtain latent features, and then extracts high-level semantic features from the latent features through a feature pyramid network. Then, the high-level semantic features are processed sequentially by two-dimensional convolution, normalization, and two-dimensional convolution to obtain high-frequency prediction coefficients. The approximation network branch first encodes the low-frequency approximation components by the encoder to obtain the encoded original features. Then, the original features are processed by three-dimensional convolution to obtain the first feature. The first feature is concatenated with the original features and then processed by two-dimensional convolution to obtain the second feature. The second feature is then processed by multiple stacked hybrid modules to obtain the third feature. Finally, the third feature is decoded by the decoder to obtain the low-frequency prediction coefficients.

5. The extreme short-term precipitation forecasting method based on wavelet domain decoupling and multi-course learning as described in claim 4, characterized in that, Each hybrid module consists of a multilayer perceptron and a residual network cascaded together.

6. The extreme short-term precipitation forecasting method based on wavelet domain decoupling and multi-course learning as described in claim 1, characterized in that, In step S2, spectral domain reconstruction is performed using discrete wavelet inverse transform: the high-frequency prediction coefficients and low-frequency approximation components are reconstructed into a first reconstruction map, the low-frequency prediction coefficients and high-frequency detail components are reconstructed into a second reconstruction map, and the low-frequency prediction coefficients and high-frequency prediction coefficients are reconstructed into a third reconstruction map.

7. The extreme short-term precipitation forecasting method based on wavelet domain decoupling and multi-course learning as described in claim 6, characterized in that, In step S2, the refining module takes the first, second, and third reconstructed images as input. The first and second reconstructed images are concatenated to form the first fused reconstructed image. The first fused reconstructed image is processed by a residual network to obtain the fourth feature. The third reconstructed image is processed by a residual network to obtain the fifth feature. The fourth feature, the fifth feature, and the second reconstructed image are concatenated to form the sixth feature. The sixth feature is processed by two-dimensional convolution to obtain the precipitation prediction map.

8. The extreme short-term precipitation forecasting method based on wavelet domain decoupling and multi-course learning as described in claim 1, characterized in that, In step S2, the loss function used during training of the prediction model consists of four parts: the first part is the regularization loss, which is the average of the first, second, and third reconstructed images; the second part is the product of the approximate branch loss and the dynamic weight scheduling function, where the approximate branch loss is the zero-mean normalized cross-correlation loss of the low-frequency prediction coefficients and the true low-frequency coefficients; the third part is the detail branch loss, which is the weighted mean square error of the high-frequency prediction coefficients and the true high-frequency coefficients; and the fourth part is the pixel domain loss, which is the mean square error between the precipitation prediction image and the true precipitation image. The true low-frequency coefficients and the true high-frequency coefficients are obtained by performing a discrete wavelet transform on the true precipitation image.

9. The extreme short-term precipitation forecasting method based on wavelet domain decoupling and multi-course learning as described in claim 8, characterized in that, The dynamic weight scheduling function takes the maximum value of two terms as the weight for approximate branch loss; the first term is the difference between 1 and the first component, which is the ratio of the current iteration round to the preset total decay steps; the second term is the preset minimum weight.

10. A wavelet domain decoupling and multi-course learning system for forecasting extreme short-term precipitation, characterized in that, include: The data processing module is used to acquire historical meteorological images containing vertical integral liquid water variables and perform preprocessing, which constitutes a meteorological image sequence containing spatiotemporal sequence information. The result acquisition module is used to input meteorological image sequences into a trained prediction model. After frequency domain decoupling of the meteorological image sequences, low-frequency approximation components and high-frequency detail components are obtained. These two components are then input into a dual-branch deep neural network composed of an approximation network branch and a detail network branch for feature extraction and prediction. The approximation network branch takes the low-frequency approximation components as input to output low-frequency prediction coefficients for future time steps; the detail network branch takes the high-frequency detail components as input to output high-frequency prediction coefficients for future time steps. Finally, spectral domain reconstruction is performed based on the low-frequency and high-frequency prediction coefficients, and the multi-path reconstruction maps are fused through a refinement module to output a precipitation prediction map. The prediction model employs a multi-task course learning strategy for end-to-end optimization, guiding the prediction model from learning the overall contour to refining local textures by dynamically adjusting the loss weights of different frequency domain components.

Images