Precipitation approach extrapolation intelligent forecasting method and device and electronic equipment

Abstract

The invention provides a rainfall proximity extrapolation intelligent forecasting method and device and electronic equipment. The method comprises the following steps: acquiring rainfall data to be forecasted; cutting the rainfall data, and inputting the cut rainfall data into the intelligent forecasting model; the intelligent forecasting model adopts a specific model structure, and comprises the step of adding a hierarchical up-sampling fusion module HUFM and a CA attention mechanism on the basis of a U-net model; performing feature extraction and deep learning rainfall extrapolation on the cut live data through an intelligent forecasting model to obtain an extrapolation forecasting result, performing adaptive weighted fusion on the extrapolation forecasting result and the forecasting data, and outputting a rainfall forecasting result in a specified time period in the future. The method is a novel deep learning nowcasting mode which gives consideration to the spatial precision, the forecasting timeliness and the rainfall intensity description capability at the same time, and can meet the urgent demands of modern meteorological services for high-resolution and intelligent rainfall prediction.

Description

Technical Field

[0001] The present application relates to the cross - technical field of artificial intelligence and meteorology, and particularly to a precipitation nowcasting intelligent prediction method, device and electronic device. Background Art

[0002] At present, precipitation nowcasting is one of the core tasks in refined meteorological services, especially playing a key role in severe convective weather warnings. Traditional extrapolation forecasting methods based on radar data, such as optical flow method, variational echo method and Lagrangian cross - correlation extrapolation method, have a certain accuracy in short - term (0 - 60 minutes) forecasting. Among them, for convective precipitation with rapid movement and evolution, the effective forecasting time is only 30 minutes, and for stratiform cloud precipitation with relatively stable movement trajectories, the forecasting time can be extended up to 120 minutes at most. However, after 2 hours, traditional radar extrapolation methods almost have no forecasting ability, especially with low forecasting accuracy for medium - to - heavy rainstorm scenarios and complex terrain areas, and the spatial resolution is not fine enough and the time extensibility is very limited.

[0003] With the development of deep learning, U - net and its improved models (such as U² - net) have been introduced into image prediction tasks and shown great potential in weather field regression prediction. Compared with the traditional U - net, U² - net introduces an innovative dual U - shaped structure ("U - in - U") in feature extraction, greatly improving its segmentation ability for fine - grained boundaries and complex structures.

[0004] However, most existing deep - learning forecasting models generally adopt a simple up - sampling method, which is difficult to restore the edge details of precipitation images and has weak modeling ability for strong precipitation areas in complex terrains, resulting in insufficient performance in the extended time of 1 - 3 hours. Summary of the Invention

[0005] The purpose of the present application is to provide a precipitation nowcasting intelligent prediction method, device and electronic device, which is a new deep - learning nowcasting method that takes into account spatial accuracy, forecasting time and precipitation intensity characterization ability at the same time, and can meet the urgent needs of modern meteorological services for high - resolution and intelligent precipitation prediction.

[0006] In a first aspect, the present application provides a precipitation nowcasting intelligent prediction method, which includes: obtaining precipitation data to be predicted; the precipitation data includes: observed data and forecast data; performing cropping processing on the precipitation data and inputting the cropped precipitation data into an intelligent prediction model; the intelligent prediction model is a model obtained by training a specific model structure with a training sample set; wherein, the training sample set includes: historical precipitation data and corresponding actual precipitation data for a specified future period; the historical precipitation data includes: precipitation analysis field data and corresponding precipitation forecast field data; the specific model structure is a model structure that adds a hierarchical upsampling fusion module HUFM and a CA attention mechanism on the basis of the U²-net model; through the U²-net model, CA attention mechanism and hierarchical upsampling fusion module HUFM in the intelligent prediction model, feature extraction and deep learning precipitation nowcasting are performed on the cropped observed data to obtain a nowcasting prediction result, and the nowcasting prediction result is adaptively weighted and fused with the forecast data to output a precipitation forecast result for a specified future period.

[0007] Further, the step of performing feature extraction and deep learning precipitation nowcasting on the cropped observed data through the U²-net model, CA attention mechanism and hierarchical upsampling fusion module HUFM in the intelligent prediction model to obtain a nowcasting prediction result includes: performing multi-level feature sampling operations on the observed data through the multi-level feature sampling U structure in the U²-net model to obtain multiple sampled features; in the order from the lowest-level feature sampling U structure to the highest-level feature sampling U structure, sequentially performing fusion processing on the sampled features through the CA attention mechanism and the hierarchical upsampling fusion module HUFM, and performing feature restoration through the multi-level feature restoration U structure to obtain multi-level sub-output features; respectively predicting the multi-level sub-output features to obtain multi-level sub-prediction results; and fusing the multi-level sub-prediction results to obtain a nowcasting prediction result.

[0008] Further, the step of sequentially performing fusion processing on the sampled features through the CA attention mechanism and the hierarchical upsampling fusion module HUFM, and performing feature restoration through the multi-level feature restoration U structure in the order from the lowest-layer feature sampling U structure to the highest-layer feature sampling U structure, and obtaining multi-level sub-output features includes: using the sampled features of the lowest-layer feature sampling U structure as the current sub-output feature of the current U structure, and performing the following feature fusion and restoration steps: processing the current sub-output feature through the hierarchical upsampling fusion module HUFM to obtain a first feature; processing the sampled features of the upper-layer U structure of the current U structure through the CA attention mechanism module to obtain a second feature; fusing the first feature and the second feature to obtain a target feature; performing feature restoration processing on the target feature through the feature restoration U structure to obtain the sub-output feature of the upper-layer U structure; using the sub-output feature of the upper-layer U structure as the current sub-output feature of the current U structure again, and continuing to perform the feature fusion and restoration steps until multi-level sub-output features are obtained.

[0009] Further, the above-mentioned hierarchical upsampling fusion module HUFM includes: a first convolutional layer, a sub-pixel convolutional layer, a SimAM attention mechanism, a second convolutional layer, a normalization layer, and a ReLU activation layer connected in sequence; the step of processing the current sub-output feature of the current U structure through the hierarchical upsampling fusion module HUFM to obtain a first feature includes: extracting local features of the current sub-output feature through the first convolutional layer; performing efficient upsampling on the local features through the sub-pixel convolutional layer to obtain upsampled features; performing feature processing on the upsampled features through the SimAM attention mechanism to obtain enhanced fusion features; the feature processing includes: feature enhancement based on 3D weights and feature fusion processing to adaptively enhance key precipitation regions and edge details; performing second convolutional processing on the enhanced fusion features through the second convolutional layer, and further optimizing the feature expression of the features after the second convolutional processing through the normalization layer and the ReLU activation layer to obtain a first feature.

[0010] Further, the training process of the above intelligent forecasting model is as follows: obtaining a training sample set; respectively cropping the precipitation analysis field data and the precipitation forecast field data in the training sample set as a frame of image; performing feature extraction and nowcasting processing on the cropped precipitation analysis field data through the U²-net model, the CA attention mechanism, and the hierarchical upsampling fusion module HUFM to obtain a deep learning nowcasting forecast result; the deep learning nowcasting forecast result includes multi-level sub-forecast results; fusing the deep learning nowcasting forecast result with the cropped precipitation forecast field data to obtain a final forecast result; calculating a loss value based on the deep learning nowcasting forecast result, the final forecast result, and the corresponding sample data, and adjusting the deep learning model parameters based on the loss value until the model converges to obtain a trained intelligent forecasting model.

[0011] Further, the step of calculating the loss value based on the deep learning extrapolation prediction result, the final prediction result, and the corresponding sample data includes: calculating the loss value according to the following formula:

[0012] wherein, represents the loss function; represents the loss between the sub-prediction result obtained by upsampling M = 6 times and the corresponding actual precipitation data in the sample; represents the loss between the final prediction result and the corresponding actual precipitation data in the sample; represents the weighted loss corresponding to the sub-prediction result after the m-th layer of sampling; represents the weighted loss corresponding to the final prediction result; , respectively represent the weight corresponding to the m-th layer of sampling and the weight of the final prediction result; is the observed value of the precipitation amount in the i-th sample, is the predicted value of the precipitation amount in the i-th sample, n is the number of samples, is the weight of the i-th sample.

[0013] Further, each of the above sample data includes: an input image sequence composed of precipitation images every 10 minutes in the recent 1 hour and predicted precipitation images in the next 1 - 3 hours, and an output image sequence including precipitation images every hour within the next 3 hours.

[0014] In a second aspect, the present application further provides a precipitation nowcasting intelligent prediction device, which includes: a data acquisition module for acquiring precipitation data to be predicted; the precipitation data includes: real-time data and forecast data; a cropping input module for cropping the precipitation data and inputting the cropped precipitation data into an intelligent prediction model; the intelligent prediction model is a model obtained by training a specific model structure with a training sample set; wherein, the training sample set includes: historical precipitation data and corresponding actual precipitation data in a specified future period; the historical precipitation data includes: precipitation analysis field data and corresponding precipitation forecast field data; the specific model structure is a model structure that adds a hierarchical upsampling fusion module HUFM and a CA attention mechanism on the basis of the U²-net model; a prediction module for performing feature extraction and deep learning precipitation extrapolation on the cropped real-time data through the U²-net model, the CA attention mechanism, and the hierarchical upsampling fusion module HUFM in the intelligent prediction model to obtain an extrapolation prediction result, and adaptively weighted fusing the extrapolation prediction result with the forecast data to output a precipitation prediction result within a specified future period.

[0015] In a third aspect, the present application further provides an electronic device, including a processor and a memory. The memory stores computer-executable instructions that can be executed by the processor, and the processor executes the computer-executable instructions to implement the method described in the above first aspect.

[0016] In a fourth aspect, the present application further provides a computer-readable storage medium. The computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are called and executed by a processor, the computer-executable instructions cause the processor to implement the method described in the above first aspect.

[0017] In the intelligent precipitation nowcasting extrapolation method, device and electronic device provided by the present application, an intelligent prediction model obtained by training a specific model structure with a training sample set is used to predict precipitation data to be predicted; wherein, the training sample set includes: historical precipitation data and corresponding actual precipitation data in a specified future period; the historical precipitation data includes: precipitation analysis field data and corresponding precipitation forecast field data; the specific model structure is a model structure that adds a hierarchical upsampling fusion module HUFM and a CA attention mechanism on the basis of the U²-net model; specifically, through the U²-net model, CA attention mechanism and hierarchical upsampling fusion module HUFM in the intelligent prediction model, feature extraction and deep learning precipitation extrapolation are performed on the cropped live data to obtain an extrapolation prediction result, and the extrapolation prediction result is adaptively weighted and fused with the forecast data to output a precipitation forecast result within a specified future period. The present application is a new deep learning nowcasting method that takes into account spatial accuracy, forecast timeliness and precipitation intensity characterization ability at the same time, and can meet the urgent needs of modern meteorological services for high-resolution and intelligent precipitation prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0019] Figure 1 It is a flowchart of an intelligent precipitation nowcasting extrapolation method provided by an embodiment of the present application; Figure 2 It is a schematic structural diagram of an intelligent prediction model provided by an embodiment of the present application; Figure 3 It is a schematic structural diagram of a hierarchical upsampling fusion module HUFM provided by an embodiment of the present application; Figure 4Schematic diagram of a CA attention mechanism module provided by an embodiment of the present application; Figure 5 Schematic diagram of a process of model training and model prediction provided by an embodiment of the present application; Figure 6 Block diagram of a structure of a precipitation nowcasting extrapolation intelligent forecasting device provided by an embodiment of the present application; Figure 7 Schematic diagram of a structure of an electronic device provided by an embodiment of the present application. Specific embodiments

[0020] Next, the technical solutions of the present application will be clearly and completely described in conjunction with the embodiments. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0021] In view of the situation that the existing precipitation nowcasting extrapolation intelligent forecasting method cannot take into account the spatial accuracy, forecasting timeliness, and precipitation intensity characterization ability, and cannot meet the urgent needs of modern meteorological services for high-resolution and intelligent precipitation prediction. Embodiments of the present application provide a precipitation nowcasting extrapolation intelligent forecasting method, device, and electronic device. By combining deep learning methods, high-resolution gridded meteorological data, and dynamic fusion strategies, a precipitation nowcasting extrapolation intelligent forecasting model is realized, and the spatial refinement and accuracy of the nowcasting precipitation forecast for the next 1 - 3 hours are improved through this model.

[0022] To facilitate the understanding of this embodiment, first, a precipitation nowcasting extrapolation intelligent forecasting method disclosed in the embodiments of the present application will be introduced in detail. Figure 1 Flowchart of a precipitation nowcasting extrapolation intelligent forecasting method provided by an embodiment of the present application. The method specifically includes the following steps: Step S102, obtain precipitation data to be forecasted; the precipitation data includes: observed data and forecast data.

[0023] The precipitation data to be forecasted described above may include: high-resolution gridded grid point meteorological data of a certain area; the observed data, that is, precipitation analysis field data, specifically refers to the gridded precipitation observed information obtained by fusing the precipitation observations of automatic weather station rain gauges and the quantitative precipitation estimation of Doppler weather radars. The precipitation analysis field data usually undergoes certain analysis and processing, and can reflect the actual precipitation conditions including precipitation amount and precipitation spatial distribution over a past period of time.

[0024] The above forecast data, i.e., precipitation forecast field data, specifically refers to the gridded precipitation forecast information obtained by simulating the atmospheric motion through vector extrapolation and numerical weather models based on a meteorological numerical fusion forecast model and forecasting future precipitation. Precipitation forecast field data is usually a prediction of the possible precipitation amount, precipitation distribution, and precipitation intensity information within a future period based on weather numerical simulation.

[0025] Step S104: Crop the precipitation data and input the cropped precipitation data into the intelligent forecast model. The intelligent forecast model is a model obtained by training a specific model structure with a training sample set. The training sample set includes historical precipitation data and the corresponding actual precipitation data for a specified future period. The historical precipitation data includes precipitation analysis field data and the corresponding precipitation forecast field data. The specific model structure is a model structure that adds a hierarchical upsampling fusion module HUFM and a CA attention mechanism to the U²-net model.

[0026] For the specific structure of the above intelligent forecast model, please refer to Figure 2 As shown, multiple hierarchical upsampling fusion modules HUFM are embedded in the conventional U²-Net double U-shaped structure to enhance the edge information recovery ability in the upsampling stage. And a CA attention mechanism module is introduced in the skip connection to capture spatial location information and improve the model's modeling ability for strong precipitation areas. The overall model structure can be regarded as an image-to-image time series regression conversion task. The specific model working process can be referred to the detailed description later.

[0027] Step S106: Through the U²-net model, CA attention mechanism, and hierarchical upsampling fusion module HUFM in the intelligent forecast model, extract features from the cropped observed data and perform deep learning precipitation extrapolation to obtain an extrapolation forecast result, and adaptively weight and fuse the extrapolation forecast result with the forecast data to output the precipitation forecast result for a specified future period.

[0028] As Figure 2 shown, the U²-net model embedded with the CA attention mechanism and hierarchical upsampling fusion module HUFM is used to extract features from the observed data and perform deep learning precipitation extrapolation to obtain an extrapolation forecast result, and then the extrapolation forecast result and the forecast data are adaptively weighted and fused to output the precipitation forecast result for a specified future period.

[0029] In the precipitation nowcasting extrapolation intelligent forecasting method provided by the embodiments of the present application, based on the dynamic information fusion of high-resolution gridded grid meteorological data, a multi-level upsampling restoration module (HUFM) and a coordinate attention mechanism (CA) are designed and introduced to construct a deep learning precipitation forecasting model. Taking a complex terrain area as an example, regional precipitation forecasting for the next 1-3 hours with a resolution of 500 meters is realized, and a forecast data product is formed. Compared with traditional extrapolation forecasting methods and general deep learning models, the deep learning intelligent forecasting method provided by this embodiment effectively extends the maximum nowcasting extrapolation precipitation forecasting time limit from 120 minutes to 180 minutes by dynamically fusing grid real-time analysis and forecast information, and significantly improves the prediction ability and model stability for heavy precipitation events, and has good potential for operational deployment.

[0030] See Figure 2 Referring to the specific structural schematic diagram of the intelligent forecasting model shown, the step of "performing feature extraction and deep learning precipitation extrapolation on the cropped real-time data through the U²-net model, CA attention mechanism, and hierarchical upsampling fusion module HUFM in the intelligent forecasting model to obtain an extrapolation forecast result" in the above step S106 specifically includes: (1) Performing multi-level feature sampling operations on the real-time data through the multi-level feature sampling U structure in the U²-net model to obtain multiple sampling features; The above multi-level feature sampling U structure is Figure 2 The multiple sub-U structures connected by the downward arrows on the left side in, and each sub-U structure outputs a sampling feature respectively. In terms of feature extraction, this model introduces the dual U-shaped structure ("U-in-U") of U²-Net to improve its segmentation ability for fine-grained boundaries and complex structures, that is, the "ReSidual U-block (RSU)" module. The RSU module adopts a nested structure similar to U-Net, further decomposes each module into a hierarchical structure of convolutional and pooling layers, and captures rich multi-scale features layer by layer through multi-level downsampling.

[0031] (2) Sequentially performing fusion processing on the sampling features through the CA attention mechanism and the hierarchical upsampling fusion module HUFM, and performing feature restoration through the multi-level feature restoration U structure in the order from the lowest-level feature sampling U structure to the highest-level feature sampling U structure to obtain multi-level sub-output features; Specifically in implementation, taking the sampling feature of the lowest-level feature sampling U structure as the current sub-output feature of the current U structure, the following feature fusion and restoration steps are performed: (2.1) Processing the current sub-output feature through the hierarchical upsampling fusion module HUFM to obtain a first feature; See Figure 3As shown, the above-mentioned hierarchical upsampling fusion module HUFM includes: a first convolutional layer, a sub-pixel convolutional layer, a SimAM attention mechanism, a second convolutional layer, a normalization layer, and a ReLU activation layer, which are connected in sequence.

[0032] When specifically implemented, the process of determining the first feature includes the following steps: (2.1.1)Extract the local features of the current sub-output feature through the first convolutional layer; (2.1.2)Perform efficient upsampling on the local features through the sub-pixel convolutional layer to obtain upsampled features; (2.1.3)Perform feature processing on the upsampled features through the SimAM attention mechanism to obtain enhanced fusion features; the feature processing includes: feature enhancement based on 3D weights and feature fusion processing to adaptively enhance the key precipitation regions and edge details; (2.1.4)Perform a second convolutional process on the enhanced fusion features through the second convolutional layer, and further optimize the feature expression of the features after the second convolutional process through the normalization layer and the ReLU activation layer to obtain the first feature.

[0033] In order to improve the upsampling ability of the model in complex heavy precipitation scenarios and enhance the recovery ability of edge structures and details, in this embodiment, based on the sub-pixel convolution and the SimAM attention mechanism, an efficient and lightweight multi-level upsampling fusion module (Hierarchical Upsampling Fusion Module, abbreviated as HUFM) is designed.

[0034] As Figure 3 shown, first, use convolution to extract local features to provide basic information for upsampling; second, use sub-pixel convolution for efficient upsampling to maintain spatial structure consistency while increasing the resolution; then, introduce the SimAM parameter-free attention mechanism to adaptively enhance the key precipitation regions and edge details without increasing the computational overhead; finally, further optimize the feature expression through convolution, batch normalization, and ReLU activation to enhance the non-linear modeling ability and alleviate the problem of gradient disappearance.

[0035] This module has good computational efficiency while ensuring accuracy, and is especially suitable for high-resolution precipitation forecasting tasks with limited resources, showing significant advantages in the edge recognition and structure restoration of heavy precipitation within 1 - 3 hours.

[0036] (2.2)Process the sampled features of the upper U structure of the current U structure through the CA attention mechanism module to obtain the second feature; In this embodiment, the CA attention mechanism module is innovatively embedded in the U-net skip connection (as Figure 2The CA attention mechanism shown) ensures that the module sizes before and after input and output remain unchanged. To enhance the model's ability to locate and recognize the structure of precipitation targets, in this embodiment, a coordinate attention mechanism (CA) is introduced in the skip connections of the model to improve the model's ability to model spatial information.

[0037] See Figure 4 The structural schematic diagram of the CA attention mechanism module shown. The CA attention mechanism module decomposes the channel attention into two one-dimensional encoding processes that are independently modeled along the spatial directions, performs global pooling along the vertical and horizontal directions respectively, enabling the model to capture long-range dependencies while retaining position information. This mechanism can effectively embed spatial coordinate information into the channel attention, enhancing the model's perception ability of complex boundaries and local precipitation structures.

[0038] While improving the feature representation ability, the CA attention mechanism module does not significantly increase the computational overhead, has good lightness and flexibility, and is suitable for accurate modeling of high-resolution precipitation features in complex scenarios.

[0039] (2.3) Perform feature fusion on the first feature and the second feature to obtain the target feature; as Figure 2 The feature output to the right by the fused character "C" in.

[0040] (2.4) Perform feature restoration processing on the target feature through the feature restoration U structure to obtain the sub-output feature of the upper U structure; The above-mentioned feature restoration U structure is Figure 2 Multiple sub-U structures symmetrically arranged with the left-side feature sampling U structure in, and gradually restore the original resolution of the feature through deconvolution, thereby outputting the sub-output feature of the upper U structure.

[0041] (2.5) Reuse the sub-output feature of the upper U structure as the current sub-output feature of the current U structure, and continue to execute the feature fusion and restoration steps until multi-level sub-output features are obtained.

[0042] From Figure 2 It can be seen that the feature restoration process is a continuously looping upward process, and the determination of the sub-output feature of the current layer depends on the sub-output feature of the upper layer. Eventually, through the loop calculation process, multi-level sub-output features can be obtained.

[0043] (3) Perform predictions on the multi-level sub-output features respectively to obtain multi-level sub-forecast results; The prediction here actually refers to the upsampling process of the sub-output features. Since the predicted data dimensions of each sub-prediction result are different, such as the output dimension of sub2 being (3, 128, 128), it needs to be upsampled to (3, 256, 256) before calculating the loss with the ground truth label (3, 256, 256) for model training.

[0044] The above-mentioned multi-level sub-forecast results such as Figure 2 sub1, sub2, sub3, sub4, sub5, sub6 shown in

[0045] (4) Fuse the multi-level sub-forecast results to obtain the extrapolation forecast result.

[0046] By fusing sub1, sub2, sub3, sub4, sub5, sub6, the extrapolation forecast result can be obtained.

[0047] See Figure 5 the schematic diagrams of the model training process and the model prediction process shown in Figure 5 which give the model training and prediction processes using data from 2021 to 2023. In the future, longer historical data can be accumulated and operated and modeled for forecasting according to the Figure 2 process. After the model is established and deployed, current real-time forecasting can also be carried out. First, the precipitation analysis field data (i.e., the ground truth analysis data) is cropped, reducing its original size (6, 1521, 1221) to (6, 256, 256) to enhance the model's ability to capture local precipitation details and also facilitate subsequent model training. Subsequently, model training is carried out. Based on the dual U-shaped structure U²-net model, a new model structure is designed (such as Figure 2 ), and then feature extraction and deep learning precipitation extrapolation are performed on the cropped data. To fully integrate the advantages of different prediction methods, the extrapolation result obtained by the deep learning method and the forecast result obtained by the traditional nowcasting cross-correlation extrapolation method ( Figure 5 precipitation forecast field data) are feature-fused through 1×1 convolution to output the final precipitation forecast result.

[0048] During the training process, the parameters of the deep learning model are optimized through continuous backpropagation, and finally a new intelligent deep learning precipitation nowcasting extrapolation model that can be stably applied to actual forecasting tasks is obtained. In the inference stage, only the precipitation data to be forecast needs to be input. Through sequential cropping and sequential input, the model can automatically complete the adaptive weighted fusion of the deep learning extrapolation result and the traditional forecast result, thereby outputting high-precision precipitation sub-forecast results (3, 256, 256) for 1 - 3 hours. Then, through sequential stitching, the final 1 - 3 hour forecast result (3, 1521, 1221) can be obtained. The training process of the above intelligent forecasting model is as follows: (1) Obtain a training sample set; In specific implementation, several typical precipitation processes in some complex areas can be selected, and the real-time and nowcasting grid data provided by the meteorological research institute can be used. Each of the above sample data includes: an input image sequence composed of precipitation images every 10 minutes in the recent 1 hour and precipitation images in the next 1 - 3 hours predicted, and an output image sequence including precipitation images every hour within the next 3 hours.

[0049] For example, in a set of training samples, the input is precipitation images every 10 minutes in the recent 1 hour, with a size of (6, 1521, 1221) + fused precipitation images in the next 1 - 3 hours predicted, with a size of (3, 1521, 1221), and the total input is (9, 1521, 1221); the output is precipitation images in the next 3 hours (every hour), with a size of (3, 1521, 1221).

[0050] (2) Consider the precipitation analysis field data and precipitation forecast field data in the training sample set as a frame of image respectively for cropping; (3) Through the U²-net model, the CA attention mechanism, and the hierarchical upsampling fusion module HUFM, perform feature extraction and nowcasting extrapolation processing on the precipitation analysis field data after image cropping to obtain the deep learning extrapolation forecast result; the deep learning extrapolation forecast result includes multi-level sub-forecast results; This process is similar to the steps in the aforementioned model prediction process, and reference can be made to the foregoing content, which will not be elaborated here.

[0051] (4) Fuse the deep learning extrapolation forecast result with the precipitation forecast field data after image cropping to obtain the final forecast result; (5) Based on the deep learning extrapolation forecast result, the final forecast result, and the corresponding sample data, calculate the loss value, and adjust the deep learning model parameters based on the loss value until the model converges to obtain a trained intelligent forecast model.

[0052] In this embodiment, the total loss of the model is calculated using the MSE (Mean Squared Error Loss) + MAE (Mean Absolute Error Loss); in specific implementation, the loss value is calculated according to the following formula:

[0053] Among them, represents the loss function; represents the loss between the sub-forecast result obtained by upsampling M = 6 times and the corresponding actual precipitation data in the sample; represents the loss between the final forecast result and the corresponding actual precipitation data in the sample; denotes the weighted loss corresponding to the sub-forecast result after sampling at the m-th layer; denotes the weighted loss corresponding to the final forecast result; , respectively denote the weight corresponding to the sampling at the m-th layer and the weight of the final forecast result; is the observed value of the precipitation in the i-th sample, is the predicted value of the precipitation in the i-th sample, and n is the number of samples, is the weight of the i-th sample.

[0054] During the model training process provided in this embodiment, the optimizer uses Adam; the batch size is adjusted to 48; the total number of iterations is set to 80; the initial learning rate is set to 0.01 and a dynamic learning rate decay strategy is introduced, that is: during the training process, if the performance index on the validation set fails to improve for three consecutive times, the learning rate will automatically decay to 50% of the original; the model training platform is PyTorch. The model input is a sequence of historical precipitation images at consecutive times, and the model output is the hourly precipitation images for the next consecutive 3 hours.

[0055] Furthermore, in this embodiment, the commonly used evaluation indicators TS (Threat Score), ETS (Equitable Threat Score), BS (Bias Score), and HSS (Heidke Skill Score) in the meteorological industry at home and abroad are also used to objectively and quantitatively evaluate the accuracy level of precipitation prediction; among them, the TS score represents the threat score, the ETS represents the equitable threat score, the BS represents the bias score, and the HSS represents the Heidke skill score. The value of the threat score ranges from 0 (completely wrong forecast) to 1 (perfect forecast), and the higher the value, the more accurate the model forecast. The value of the equitable threat score also ranges from 0 to 1, and the higher the value, the higher the skill of the forecast, and the bias caused by random chance is eliminated. The closer the value of the bias score is to 1, the smaller the bias of the model. Greater than (less than) 1 indicates that the prediction intensity is too strong (too weak), the range is too large (too small), and it is easy to have false alarms (missed alarms). The value of the Heidke skill score is between -1 and 1. 0 indicates that the forecast skill of the model is equivalent to a random forecast, 1 indicates perfect forecast skill, and a negative value indicates that the model is worse than a random forecast. It is a comprehensive indicator for evaluating the skill of a forecast model.

[0056] Under the condition of a precipitation threshold of more than 10 mm / h, the method and its prediction model provided in this embodiment have the highest improvement of 52.02% in TS and ETS scores and 22.22% in HSS compared with the existing short-term precipitation prediction level. The BS is closer to the optimal value of 1. Compared with the traditional cross-correlation extrapolation method and the U-net and U²-net baseline deep learning models, the method and its prediction model provided in this embodiment have achieved significant performance improvements, and the effective prediction time of precipitation nowcasting has been extended from 120 minutes to 180 minutes.

[0057] Different from the traditional deep learning precipitation extrapolation model, after generating the preliminary extrapolation result in this embodiment, it performs 1×1 convolution dynamic fusion with the 500-meter resolution numerical fusion prediction field, and uses the fused precipitation prediction map as the feedback input to participate in the subsequent training and prediction processes, forming a closed-loop optimization mechanism. The advantages of this feedback mechanism are as follows: (1) Enhance the prediction accuracy and capture multi-scale information: Fusing the numerical prediction results enables the model to not only have the ability to capture strong precipitation areas but also obtain the simulation advantages of the numerical model for large-scale precipitation systems, improving the overall spatio-temporal consistency and prediction accuracy.

[0058] (2) Improve the model robustness and reduce the bias of a single model: Mitigate the problem of model error accumulation under complex weather through the information complementary mechanism, and improve the adaptability to various precipitation types (convective, stratiform clouds).

[0059] (3) Extend the extrapolation time: The prediction skill of the traditional extrapolation model decays significantly after 60 minutes and almost completely loses its prediction skill after 120 minutes. However, fusing the numerical prediction results can effectively alleviate this problem and extend the effective prediction time from 120 minutes to 180 minutes.

[0060] The precipitation nowcasting intelligent prediction method provided in the embodiment of this application has the following remarkable advantages: 1. Achieve 1-3 hour nowcasting of deep learning precipitation with a 500-meter high resolution based on grid meteorological data, leading in the country; [[ID=~18]]2. The HUFM module significantly improves the image upsampling effect and enhances the prediction of local strong precipitation by introducing sub-pixel convolution and SimAM attention mechanism; 3. The CA mechanism strengthens the model's ability to model spatial context relationships and improves the accuracy of precipitation area positioning; 4. By dynamically fusing numerical prediction information, effectively extend the precipitation prediction time to 180 minutes, significantly improve the prediction ability and model stability for strong precipitation events, and have good potential for operational deployment.

[0061] Based on the above method embodiment, the embodiment of this application also provides a precipitation nowcasting intelligent prediction device. SeeFigure 6 As shown, the device includes: a data acquisition module 62 for acquiring precipitation data to be predicted; the precipitation data includes: actual data and forecast data; a cropping input module 64 for cropping the precipitation data and inputting the cropped precipitation data into an intelligent prediction model; the intelligent prediction model is a model obtained by training a specific model structure with a training sample set; wherein, the training sample set includes: historical precipitation data and corresponding actual precipitation data for a specified future period; the historical precipitation data includes: precipitation analysis field data and corresponding precipitation forecast field data; the specific model structure is a model structure that adds a hierarchical upsampling fusion module HUFM and a CA attention mechanism on the basis of the U²-net model; a prediction module 66 for performing feature extraction and deep learning precipitation extrapolation on the cropped actual data through the U²-net model, CA attention mechanism and hierarchical upsampling fusion module HUFM in the intelligent prediction model, obtaining an extrapolation prediction result, and adaptively weighted fusing the extrapolation prediction result with the forecast data to output a precipitation prediction result for a specified future period.

[0062] Further, the above prediction module 66 is used to perform multi-level feature sampling operations on the actual data through the multi-level feature sampling U structure in the U²-net model to obtain multiple sampling features; in the order from the lowest-level feature sampling U structure to the highest-level feature sampling U structure, the sampling features are successively fused through the CA attention mechanism and the hierarchical upsampling fusion module HUFM, and feature restoration is performed through the multi-level feature restoration U structure to obtain multi-level sub-output features; the multi-level sub-output features are respectively predicted to obtain multi-level sub-prediction results; the multi-level sub-prediction results are fused to obtain an extrapolation prediction result.

[0063] Further, the above prediction module 66 is used to use the sampling feature of the lowest-level feature sampling U structure as the current sub-output feature of the current U structure and perform the following feature fusion and restoration steps: process the current sub-output feature through the hierarchical upsampling fusion module HUFM to obtain a first feature; process the sampling feature of the upper-level U structure of the current U structure through the CA attention mechanism module to obtain a second feature; fuse the first feature and the second feature to obtain a target feature; perform feature restoration processing on the target feature through the feature restoration U structure to obtain the sub-output feature of the upper-level U structure; use the sub-output feature of the upper-level U structure as the current sub-output feature of the current U structure again and continue to perform the feature fusion and restoration steps until multi-level sub-output features are obtained.

[0064] Further, the above-mentioned hierarchical upsampling fusion module HUFM includes: a first convolutional layer, a sub-pixel convolutional layer, a SimAM attention mechanism, a second convolutional layer, a normalization layer, and a ReLU activation layer connected in sequence; a prediction module 66, which is used to extract local features of the current sub-output features through the first convolutional layer; perform efficient upsampling on the local features through the sub-pixel convolutional layer to obtain upsampled features; perform feature processing on the upsampled features through the SimAM attention mechanism to obtain enhanced fusion features; the feature processing includes: feature enhancement and feature fusion processing based on 3D weights to adaptively enhance key precipitation regions and edge details; perform a second convolutional process on the enhanced fusion features through the second convolutional layer, and further optimize the feature expression of the features after the second convolutional process through the normalization layer and the ReLU activation layer to obtain the first feature.

[0065] Further, the above-mentioned device further includes: a model training module, which is used to execute the following training process of the intelligent prediction model: obtain a training sample set; respectively crop the precipitation analysis field data and the precipitation forecast field data in the training sample set as one frame of image; through the U²-net model, the CA attention mechanism, and the hierarchical upsampling fusion module HUFM, perform feature extraction and nowcasting extrapolation processing on the precipitation analysis field data after image cropping to obtain a deep learning extrapolation forecast result; the deep learning extrapolation forecast result includes multi-level sub-forecast results; fuse the deep learning extrapolation forecast result with the precipitation forecast field data after image cropping to obtain a final forecast result; calculate a loss value based on the deep learning extrapolation forecast result, the final forecast result, and the corresponding sample data, and adjust the deep learning model parameters based on the loss value until the model converges to obtain a trained intelligent prediction model.

[0066] Further, the above-mentioned model training module is used to calculate the loss value according to the following formula:

[0067] where, represents the loss function; represents the loss between the sub-forecast result obtained by upsampling M = 6 times and the corresponding actual precipitation data in the sample; represents the loss between the final forecast result and the corresponding actual precipitation data in the sample; represents the weighted loss corresponding to the sub-forecast result after the m-th sampling; represents the weighted loss corresponding to the final forecast result; , respectively represent the weight corresponding to the m-th sampling, and the weight of the final forecast result; is the observed value of the precipitation amount in the i-th sample, is the predicted value of the precipitation amount in the i-th sample, and n is the number of samples. is the weight of the i-th sample.

[0068] Furthermore, each of the above sample data includes: an input image sequence composed of precipitation images every 10 minutes in the recent 1 hour and precipitation images predicted for the next 1 - 3 hours, and an output image sequence including precipitation images every hour within the next 3 hours.

[0069] The device provided by the embodiment of the present application has the same implementation principle and technical effects as those of the foregoing method embodiment. For the sake of brief description, for the parts not mentioned in the device embodiment, reference may be made to the corresponding content in the foregoing method embodiment. <S

[0070] The embodiment of the present application also provides an electronic device, as Figure 7 shown, which is a schematic structural diagram of the electronic device. Among them, the electronic device includes a processor 71 and a memory 70. The memory 70 stores computer-executable instructions that can be executed by the processor 71, and the processor 71 executes the computer-executable instructions to implement the above method.

[0071] In Figure 7 the shown embodiment, the electronic device further includes a bus 72 and a communication interface 73. Among them, the processor 71, the communication interface 73, and the memory 70 are connected through the bus 72.

[0072] Among them, the memory 70 may include a high-speed random access memory (RAM, Random Access Memory), and may also include a non-volatile memory, such as at least one disk memory. Through at least one communication interface 73 (which can be wired or wireless), a communication connection is realized between the system network element and at least one other network element, and the Internet, wide area network, local area network, metropolitan area network, etc. can be used. The bus 72 may be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus 72 can be divided into an address bus, a data bus, a control bus, etc. For the sake of convenience of representation, Figure 7 only a bidirectional arrow is used in

[0073] The processor 71 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit in the hardware of the processor 71 or instructions in the form of software. The above-mentioned processor 71 may be a general-purpose processor, including a Graphic Processing Unit (GPU), a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the method disclosed in the embodiments of the present application can be directly implemented by the hardware decoding processor, or implemented by a combination of the hardware and software modules in the decoding processor. The software module may be located in a mature storage medium in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. This storage medium is located in the memory, and the processor 71 reads the information in the memory and combines its hardware to complete the steps of the method in the foregoing embodiments.

[0074] The embodiments of the present application also provide a computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are called and executed by the processor, the computer-executable instructions cause the processor to implement the above method. For specific implementation, reference may be made to the foregoing method embodiments, and details are not described herein again.

[0075] The computer program product of the method, apparatus, and electronic device provided by the embodiments of the present application includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the method described in the foregoing method embodiments. For specific implementation, reference may be made to the method embodiments, and details are not described herein again.

[0076] Unless otherwise specifically stated, the relative steps, numerical expressions, and values of the components and steps set forth in these embodiments do not limit the scope of the present application. [[ID= 10]]

[0077] When the above-mentioned functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium executable by a processor. Based on such understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of this 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 enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of this application. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs.

[0078] In the description of this application, it should be noted that the orientation or positional relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing this application and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it should not be understood as a limitation to this application. In addition, the terms "first", "second", "third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0079] Finally, it should be noted that: the above-mentioned embodiments are only specific implementation manners of this application, used to illustrate the technical solutions of this application, rather than limiting it. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: any person skilled in the art within the technical scope disclosed in this application can still modify the technical solutions recorded in the foregoing embodiments, or can easily think of changes, or perform equivalent replacements on some of the technical features; and these modifications, changes, or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be subject to the protection scope of the claims.

Claims

1. An intelligent prediction method for precipitation nowcasting extrapolation, characterized in that The method includes: Obtaining precipitation data to be predicted; the precipitation data includes: observed data and forecast data; Performing cropping processing on the precipitation data, and inputting the cropped precipitation data into an intelligent forecasting model; the intelligent forecasting model is a model obtained by training a specific model structure with a training sample set; wherein, the training sample set includes: historical precipitation data and corresponding actual precipitation data for a specified future period; the historical precipitation data includes: precipitation analysis field data and corresponding precipitation forecast field data; the specific model structure is a model structure that adds a hierarchical upsampling fusion module HUFM and a CA attention mechanism on the basis of the U²-net model; Through the U²-net model, CA attention mechanism and the hierarchical upsampling fusion module HUFM in the intelligent forecasting model, performing feature extraction and deep learning precipitation extrapolation on the cropped observed data to obtain an extrapolation forecast result, and adaptively weighted fusing the extrapolation forecast result with the forecast data to output a precipitation forecast result for a specified future period.

2. The method according to claim 1, wherein The step of performing feature extraction and deep learning precipitation extrapolation on the cropped observed data through the U²-net model, CA attention mechanism and the hierarchical upsampling fusion module HUFM in the intelligent forecasting model to obtain an extrapolation forecast result includes: Performing multi-level feature sampling operations on the observed data through the multi-level feature sampling U structure in the U²-net model to obtain multiple sampled features; In the order from the bottommost feature sampling U structure to the topmost feature sampling U structure, successively performing fusion processing on the sampled features through the CA attention mechanism and the hierarchical upsampling fusion module HUFM, and performing feature restoration through the multi-level feature restoration U structure to obtain multi-level sub-output features; Respectively predicting the multi-level sub-output features to obtain multi-level sub-forecast results; Fusing the multi-level sub-forecast results to obtain an extrapolation forecast result.

3. The method according to claim 2, characterized in that The step of successively performing fusion processing on the sampled features through the CA attention mechanism and the hierarchical upsampling fusion module HUFM, and performing feature restoration through the multi-level feature restoration U structure in the order from the bottommost feature sampling U structure to the topmost feature sampling U structure to obtain multi-level sub-output features includes: Taking the sampled feature of the bottommost feature sampling U structure as the current sub-output feature of the current U structure, and performing the following feature fusion and restoration steps: Processing the current sub-output feature through the hierarchical upsampling fusion module HUFM to obtain a first feature; Processing the sampled feature of the upper-layer U structure of the current U structure through the CA attention mechanism module to obtain a second feature; Performing feature fusion on the first feature and the second feature to obtain a target feature; Feature restoration processing of the target feature is performed through the U structure with feature restoration to obtain the sub-output feature of the upper-layer U structure; the sub-output feature of the upper-layer U structure is re-used as the current sub-output feature of the current U structure, and the feature fusion and restoration steps are continued until multi-level sub-output features are obtained.

4. The method according to claim 3, wherein The hierarchical up-sampling fusion module HUFM includes: a first convolutional layer, a sub-pixel convolutional layer, a SimAM attention mechanism, a second convolutional layer, a normalization layer, and a ReLU activation layer connected in sequence; the step of processing the current sub-output feature of the current U structure through the hierarchical up-sampling fusion module HUFM to obtain the first feature includes: Extracting local features of the current sub-output feature through the first convolutional layer; Performing efficient up-sampling on the local features through the sub-pixel convolutional layer to obtain up-sampled features; Performing feature processing on the up-sampled features through the SimAM attention mechanism to obtain enhanced fusion features; the feature processing includes: feature enhancement and feature fusion processing based on 3D weights to adaptively enhance key precipitation regions and edge details; Performing second convolutional processing on the enhanced fusion features through the second convolutional layer, and further optimizing the feature expression of the features after the second convolutional processing through the normalization layer and the ReLU activation layer to obtain the first feature.

5. The method according to claim 1, characterized in that The training process of the intelligent prediction model is as follows: Obtaining a training sample set; Cropping the precipitation analysis field data and the precipitation forecast field data in the training sample set as a frame of image respectively; Performing feature extraction and nowcasting extrapolation processing on the precipitation analysis field data after image cropping through the U²-net model, the CA attention mechanism, and the hierarchical up-sampling fusion module HUFM to obtain a deep learning extrapolation forecast result; the deep learning extrapolation forecast result includes multi-level sub-forecast results; Fusing the deep learning extrapolation forecast result with the precipitation forecast field data after image cropping to obtain a final forecast result; Calculating a loss value based on the deep learning extrapolation forecast result, the final forecast result, and the corresponding sample data, and adjusting the deep learning model parameters based on the loss value until the model converges to obtain a trained intelligent prediction model.

6. The method according to claim 5, characterized in that, The step of calculating the loss value based on the deep learning extrapolation forecast result, the final forecast result, and the corresponding sample data includes: Calculating the loss value according to the following formula: Among them, represents the loss function; represents the loss between the sub-forecast result obtained by upsampling M = 6 times and the corresponding actual precipitation data in the sample; represents the loss between the final forecast result and the corresponding actual precipitation data in the sample; represents the weighted loss corresponding to the sub-forecast result after sampling at the m-th layer; represents the weighted loss corresponding to the final forecast result; and represent the weight corresponding to sampling at the m-th layer and the weight of the final forecast result, respectively; is the observed value of the precipitation amount in the i-th sample, is the predicted value of the precipitation amount in the i-th sample, n is the number of samples, is the weight of the i-th sample.

7. The method according to claim 5, wherein Each sample data includes: an input image sequence composed of precipitation images every 10 minutes in the past 1 hour and precipitation images in the future 1-3 hours to be forecasted, and an output image sequence including precipitation images every hour within the next 3 hours.

8. An intelligent nowcasting device for precipitation extrapolation, characterized in that, The device includes: A data acquisition module for acquiring precipitation data to be forecasted; the precipitation data includes: real-time data and forecast data; A cropping input module for cropping the precipitation data and inputting the cropped precipitation data into an intelligent forecasting model; the intelligent forecasting model is a model obtained by training a specific model structure with a training sample set; wherein, the training sample set includes: historical precipitation data and corresponding actual precipitation data for a specified future period; the historical precipitation data includes: precipitation analysis field data and corresponding precipitation forecast field data; the specific model structure is a model structure that adds a hierarchical upsampling fusion module HUFM and a CA attention mechanism on the basis of the U²-net model. A forecasting module for performing feature extraction and deep learning precipitation extrapolation on the cropped live data through the U²-net model, CA attention mechanism and the hierarchical upsampling fusion module HUFM in the intelligent forecasting model, obtaining an extrapolation forecast result, and adaptively weighted fusing the extrapolation forecast result with the forecast data to output a precipitation forecast result for a specified future period.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory stores computer executable instructions that can be executed by the processor, and the processor executes the computer executable instructions to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer executable instructions, and when the computer executable instructions are called and executed by the processor, the computer executable instructions cause the processor to implement the method according to any one of claims 1 to 7.

Images