Precipitation near extrapolation intelligent prediction method and device and electronic equipment

Abstract

The application provides a precipitation near extrapolation intelligent prediction method and device and electronic equipment, and the method comprises the following steps: obtaining precipitation data to be predicted; performing clipping processing on the precipitation data, and inputting the clipped precipitation data into an intelligent prediction model; the intelligent prediction model adopts a specific model structure, which comprises: adding a hierarchical upsampling fusion module HUFM and a CA attention mechanism on the basis of a U2-net model; through the intelligent prediction model, feature extraction and deep learning precipitation extrapolation are performed on the clipped real-time data, an extrapolation prediction result is obtained, the extrapolation prediction result is adaptively weighted with the prediction data, and a precipitation prediction result in a specified period in the future is output. The application is a new type of deep learning nowcasting method which simultaneously considers spatial accuracy, prediction timeliness and precipitation intensity description capability, and can meet the urgent needs of modern meteorological services for high-resolution and intelligent precipitation prediction.

Description

Technical Field

[0001] The present application relates to the interdisciplinary field of artificial intelligence and meteorology, and in particular to a precipitation nowcasting and extrapolation intelligent forecasting method, device, and electronic equipment. Background Art

[0002] Currently, precipitation nowcasting is a core task in refined meteorological services, playing a particularly crucial role in severe convective weather warnings. Traditional radar data-based extrapolation forecasting methods, such as optical flow, variational echo, and Lagrangian cross-correlation, offer some accuracy in near-term forecasts (0–60 minutes). However, for rapidly evolving convective precipitation, the effective forecast timeframe is only 30 minutes, while for relatively stable stratiform clouds, the forecast timeframe can be extended to 120 minutes. However, beyond two hours, traditional radar extrapolation methods have little predictive capability, particularly for moderate to heavy rainfall scenarios and areas with complex terrain. Furthermore, their accuracy is low, and their spatial resolution is not sufficiently refined, and their temporal scalability is very limited.

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

[0004] However, most existing deep learning forecasting models generally use simple upsampling methods, which make it difficult to restore the edge details of precipitation images, and have weak modeling capabilities for heavy precipitation areas in complex terrain, resulting in insufficient performance under extended time constraints of 1-3 hours. Summary of the Invention

[0005] The purpose of this application is to provide a precipitation nowcasting intelligent forecasting method, device and electronic equipment, which is a new deep learning nowcasting method that takes into account spatial accuracy, forecast timeliness and precipitation intensity characterization capabilities. It can meet the urgent needs of modern meteorological services for high-resolution and intelligent precipitation forecasting.

[0006] In the first aspect, the present application provides a method for intelligent forecasting of precipitation near-term extrapolation, the method comprising: obtaining precipitation data to be forecasted; the precipitation data comprising: actual data and forecast data; 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 through a training sample set; wherein the training sample set comprises: historical precipitation data and the corresponding actual precipitation data for a specified future period; the historical precipitation data comprises: precipitation analysis field data and the corresponding precipitation forecast field data; the specific model structure is a model structure in which a hierarchical upsampling fusion module HUFM and a CA attention mechanism are added to the U²-net model; through the U²-net model, the CA attention mechanism and the hierarchical upsampling fusion module HUFM in the intelligent forecasting model, feature extraction and deep learning precipitation extrapolation are performed on the cropped actual data to obtain an extrapolated forecast result, and the extrapolated forecast result is adaptively weighted fused with the forecast data to output the precipitation forecast result for the specified future period.

[0007] Furthermore, the above-mentioned step of 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 forecast model to obtain the extrapolated forecast result includes: performing 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; fusing the sampling features through the CA attention mechanism and the hierarchical upsampling fusion module HUFM in order from the lowest feature sampling U structure to the highest feature sampling U structure, and performing feature recovery through the multi-level feature recovery U structure to obtain multi-level sub-output features; predicting the multi-level sub-output features respectively to obtain multi-level sub-forecast results; and fusing the multi-level sub-forecast results to obtain the extrapolated forecast result.

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

[0009] Further, the hierarchical upsampling fusion module HUFM comprises: 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 features of the current U structure through the hierarchical upsampling fusion module HUFM to obtain the first features comprises: extracting local features of the current sub-output features 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 comprises feature enhancement and feature fusion processing based on 3D weights to adaptively enhance key precipitation areas and edge details; performing second convolution processing on the enhanced fusion features through the second convolutional layer, and further optimizing the feature expression of the features after the second convolution processing through the normalization layer and the ReLU activation layer to obtain the first features.

[0010] Further, 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 prediction field data in the training sample set as a frame of image respectively; performing feature extraction and nearby extrapolation 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 extrapolation prediction result; the deep learning extrapolation prediction result comprises multi-level sub-prediction results; fusing the deep learning extrapolation prediction result with the cropped precipitation prediction field data to obtain a final prediction result; calculating a loss value based on the deep learning extrapolation prediction result, the final prediction result, and the corresponding sample data, and adjusting the deep learning model parameters based on the loss value until the model converges to obtain the trained intelligent prediction model.

[0011] Furthermore, the step of calculating the loss value based on the deep learning extrapolated forecast results, the final forecast results and the corresponding sample data includes: calculating the loss value according to the following formula:

[0012]

[0013] in, 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; It 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-prediction result after sampling at the mth layer; Represents the weighted loss corresponding to the final forecast result; 、 They represent the weight corresponding to the m-th layer sampling and the weight of the final forecast result respectively; is the observed value of precipitation in the i-th sample, is the predicted value of precipitation in the i-th sample, n is the number of samples, is the weight of the i-th sample.

[0014] Furthermore, each of the sample data includes: an input image sequence consisting of 10-minute precipitation images for the past hour and predicted precipitation images for the next 1-3 hours, and an output image sequence including hourly precipitation images for the next 3 hours.

[0015] In the second aspect, the present application also provides an intelligent forecasting device for precipitation extrapolation, the device including: a data acquisition module for acquiring precipitation data to be forecasted; the precipitation data includes: actual data and forecast data; a clipping input module for clipping the precipitation data and inputting the clipped precipitation data into the intelligent forecasting model; the intelligent forecasting model is a model obtained by training a specific model structure through a training sample set; wherein 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 in which a layered upsampling fusion module HUFM and a CA attention mechanism are added to the U²-net model; a forecasting module is used to perform feature extraction and deep learning precipitation extrapolation on the clipped actual data through the U²-net model, CA attention mechanism and layered upsampling fusion module HUFM in the intelligent forecasting model to obtain an extrapolated forecast result, and adaptively weighted fuse the extrapolated forecast result with the forecast data to output the precipitation forecast result for the specified future period.

[0016] In a third aspect, the present application further provides an electronic device comprising a processor and a memory, wherein 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 first aspect above.

[0017] In a fourth aspect, the present application also provides a computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions prompt the processor to implement the method described in the first aspect above.

[0018] In the precipitation near-term extrapolation intelligent forecasting method, device, and electronic device provided in the present application, the intelligent forecasting model obtained by training a specific model structure with a training sample set predicts the precipitation data to be forecasted; wherein, 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 with a layered upsampling fusion module HUFM and a CA attention mechanism added to the U²-net model; specifically, through the U²-net model, CA attention mechanism, and layered upsampling fusion module HUFM in the intelligent forecasting model, feature extraction and deep learning precipitation extrapolation are performed on the cropped actual data to obtain an extrapolated forecast result, and the extrapolated forecast result is adaptively weighted fused with the forecast data to output the precipitation forecast result for the specified future period. The present application is a new deep learning near-term forecasting method that takes into account spatial accuracy, forecast timeliness, and precipitation intensity characterization capabilities, and can meet the urgent needs of modern meteorological services for high-resolution, intelligent precipitation forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the specific implementation methods of the present application or the technical solutions in the prior art, the following is a brief introduction to the drawings required for use in the specific implementation methods or the description of the prior art. Obviously, the drawings described below are some implementation methods of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0020] Figure 1 A flowchart of a precipitation nowcasting and extrapolation intelligent forecasting method provided in an embodiment of the present application;

[0021] Figure 2 A schematic diagram of the structure of an intelligent forecasting model provided in an embodiment of the present application;

[0022] Figure 3 A schematic diagram of the structure of a hierarchical upsampling fusion module HUFM provided in an embodiment of the present application;

[0023] Figure 4 A schematic diagram of the structure of a CA attention mechanism module provided in an embodiment of the present application;

[0024] Figure 5 A schematic diagram of a model training and model prediction process provided in an embodiment of the present application;

[0025] Figure 6 This is a structural block diagram of a precipitation nowcasting and extrapolation intelligent forecasting device provided in an embodiment of the present application;

[0026] Figure 7 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0027] The following will clearly and completely describe the technical solutions of this application in conjunction with the embodiments. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0028] Existing precipitation near-field extrapolation intelligent forecasting methods are unable to balance spatial accuracy, forecast timeliness, and precipitation intensity characterization capabilities, and cannot meet the urgent needs of modern meteorological services for high-resolution, intelligent precipitation forecasting. The present application provides a precipitation near-field extrapolation intelligent forecasting method, device, and electronic device. By combining deep learning methods, high-resolution gridded meteorological data, and dynamic fusion strategies, a precipitation near-field extrapolation intelligent forecasting model is implemented. This model improves the spatial refinement and accuracy of the precipitation near-field forecast for the next 1-3 hours.

[0029] To facilitate understanding of this embodiment, a precipitation near-falling extrapolation intelligent forecasting method disclosed in an embodiment of the present application is first introduced in detail.

[0030] Figure 1 This is a flowchart of a precipitation nowcasting intelligent forecasting method provided in an embodiment of the present application, which specifically includes the following steps:

[0031] Step S102: Obtain precipitation data to be forecasted; the precipitation data includes: actual data and forecast data.

[0032] The aforementioned precipitation data to be forecasted can include high-resolution gridded meteorological data for a specific region; the aforementioned real-time data, also known as precipitation analysis field data, refers to gridded real-time precipitation information derived from a fusion of precipitation observations from rain gauges at automatic weather stations and quantitative precipitation estimates from Doppler weather radar. Precipitation analysis field data typically undergoes certain analysis and processing to reflect actual precipitation conditions over a period of time, including precipitation amount and spatial distribution.

[0033] This forecast data, also known as precipitation forecast field data, refers to gridded precipitation forecast information derived from a meteorological numerical fusion forecast model, which simulates atmospheric motion through vector extrapolation and numerical weather models, and predicts future precipitation conditions. Precipitation forecast field data typically provides predictions of the likely amount, distribution, and intensity of precipitation over a specific period of time, based on numerical weather simulations.

[0034] In step S104, the precipitation data is cropped and input into the intelligent forecasting model. The intelligent forecasting model is a model obtained by training a specific model structure using a training sample set. The training sample set includes historical precipitation data and 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 U²-net model with the addition of a hierarchical upsampling fusion module (HUFM) and a CA attention mechanism.

[0035] The specific structure of the above intelligent forecast model can be found in Figure 2 As shown in the figure, multiple hierarchical upsampling fusion modules (HUFMs) are embedded in the conventional U²-Net dual U-shaped structure to enhance the edge information recovery capability in the upsampling stage. Furthermore, a CA attention mechanism module is introduced in the skip connection to capture spatial location information and improve the model's ability to model heavy rainfall areas. The overall model structure can be viewed as an image-to-image temporal regression conversion task. The specific model working process can be found in the detailed description below.

[0036] In step S106, the U²-net model, CA attention mechanism, and layered upsampling fusion module HUFM in the intelligent forecast model are used to perform feature extraction and deep learning precipitation extrapolation on the cropped live data to obtain an extrapolated forecast result. The extrapolated forecast result is then adaptively weighted fused with the forecast data to output the precipitation forecast result for the specified future period.

[0037] like Figure 2As shown in the , a U²-net model embedded with the CA attention mechanism and the hierarchical upsampling fusion module HUFM is used to perform feature extraction and deep learning precipitation extrapolation on the actual data to obtain the extrapolated forecast results. Then, the extrapolated forecast results are adaptively weighted fused with the forecast data to output the precipitation forecast results for a specified period in the future.

[0038] The precipitation nowcasting intelligent forecasting method provided in this embodiment of the application utilizes dynamic information fusion of high-resolution gridded meteorological data, designs and introduces a multi-level upsampling recovery module (HUFM) and a coordinate attention mechanism (CA) to construct a deep learning precipitation forecast model. Taking complex terrain areas as an example, this method achieves a 1-3 hour regional precipitation forecast with a 500-meter resolution, generating a forecast data product. Compared to traditional extrapolation forecasting methods and general deep learning models, the deep learning intelligent forecasting method provided in this embodiment effectively extends the maximum timeliness of radar nowcast precipitation forecasts from 120 minutes to 180 minutes by dynamically fusing grid point real-time analysis and forecast information. It also significantly improves the predictive capability and model stability for heavy precipitation events, demonstrating excellent potential for commercial deployment.

[0039] See also Figure 2 The specific structural diagram of the intelligent forecast model shown in the figure, the step of "using the U²-net model, CA attention mechanism and layered upsampling fusion module HUFM in the intelligent forecast model to perform feature extraction and deep learning precipitation extrapolation on the cropped real-time data to obtain the extrapolated forecast result" in the above step S106 specifically includes:

[0040] (1) Through the multi-level feature sampling U structure in the U²-net model, multi-level feature sampling operations are performed on the real data to obtain multiple sampling features;

[0041] The above multi-level feature sampling U structure is Figure 2 The downward arrows on the left side of the figure connect multiple sub-U structures, each of which outputs a sampled feature. This model incorporates the dual U-shaped structure ("U-in-U") of U²-Net for feature extraction, improving its ability to segment fine-grained boundaries and complex structures. This is known as the "ReSidual U-block (RSU)" module. The RSU module employs a nested structure similar to U-Net, further decomposing each module into a hierarchy of convolutional and pooling layers. Through multi-level downsampling, it captures rich multi-scale features layer by layer.

[0042] (2) In the order from the lowest feature sampling U structure to the highest feature sampling U structure, the sampled features are fused through the CA attention mechanism and the hierarchical upsampling fusion module HUFM, and the features are restored through the multi-level feature recovery U structure to obtain multi-level sub-output features;

[0043] In specific implementation, the sampling features of the bottom-level feature sampling U structure are used as the current sub-output features of the current U structure, and the following feature fusion and recovery steps are performed:

[0044] (2.1) The current sub-output feature is processed through the hierarchical upsampling fusion module HUFM to obtain the first feature;

[0045] See also Figure 3 As shown in the figure, 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.

[0046] In specific implementation, the process of determining the first feature includes the following steps:

[0047] (2.1.1) Extract local features of the current sub-output features through the first convolutional layer;

[0048] (2.1.2) Efficiently upsample local features through sub-pixel convolutional layers to obtain upsampled features;

[0049] (2.1.3) Feature processing is performed on the upsampled features using the SimAM attention mechanism to obtain enhanced fusion features. Feature processing includes: 3D weight-based feature enhancement and feature fusion processing to adaptively enhance key precipitation areas and edge details.

[0050] (2.1.4) The enhanced fusion features are subjected to a second convolution process through the second convolution layer, and the feature expression of the features after the second convolution process is further optimized through the normalization layer and the ReLU activation layer to obtain the first feature.

[0051] To improve the model's upsampling capability in complex heavy rainfall scenarios and enhance its ability to recover edge structures and details, this embodiment designs an efficient and lightweight Hierarchical Upsampling Fusion Module (HUFM) based on sub-pixel convolution and the SimAM attention mechanism.

[0052] like Figure 3 As shown in , first, convolution is used to extract local features to provide basic information for upsampling; second, sub-pixel convolution is used for efficient upsampling to maintain spatial structure consistency while improving resolution; then, the SimAM parameter-free attention mechanism is introduced to adaptively enhance key precipitation areas and edge details without increasing computational overhead; finally, convolution, batch normalization and ReLU activation are used to further optimize feature expression, enhance nonlinear modeling capabilities, and alleviate the gradient vanishing problem.

[0053] This module has good computational efficiency while ensuring accuracy, and is particularly suitable for resource-constrained high-resolution precipitation forecast tasks. It shows significant advantages in edge recognition and structure restoration of heavy rainfall within 1-3 hours.

[0054] (2.2) The sampling features of the previous U structure of the current U structure are processed through the CA attention mechanism module to obtain the second feature;

[0055] In this embodiment, the CA attention mechanism module is innovatively embedded into the U-net jump connection (e.g. Figure 2 The CA attention mechanism shown in the figure ensures that the module size remains unchanged before and after input and output. To enhance the model's ability to locate and identify the structure of precipitation targets, this embodiment introduces a coordinate attention mechanism (CA) in the model's skip connections to improve the model's ability to model spatial information.

[0056] See also Figure 4 The CA attention mechanism module is shown in Figure 2. The module decomposes channel attention into two independently modeled one-dimensional encoding processes along the spatial direction, performing global pooling in the vertical and horizontal directions. This allows the model to capture long-range dependencies while preserving position information. This mechanism effectively embeds spatial coordinate information into channel attention, improving the model's perception of complex boundaries and local precipitation structures.

[0057] The CA attention mechanism module improves the feature representation capability without significantly increasing the computational overhead. It has good lightness and flexibility and is suitable for accurate modeling of high-resolution precipitation features in complex scenarios.

[0058] (2.3) Fusing the first feature and the second feature to obtain the target feature; Figure 2 The features of the character "C" output to the right are fused.

[0059] (2.4) Perform feature recovery processing on the target feature through the feature recovery U structure to obtain the sub-output features of the previous layer U structure;

[0060] The above characteristics restore the U structure to Figure 2 The multiple sub-U structures in the middle are symmetrically arranged with the feature sampling U structure on the left. The original resolution of the features is gradually restored through deconvolution, thereby outputting the sub-output features of the U structure on the previous layer.

[0061] (2.5) The sub-output features of the previous U-structure are used as the current sub-output features of the current U-structure, and the feature fusion and recovery steps are continued until multi-level sub-output features are obtained.

[0062] from Figure 2As can be seen from the figure, the feature recovery process is a continuous upward process. The determination of the sub-output features of the current layer depends on the sub-output features of the previous layer. Finally, through the cyclic calculation process, multi-level sub-output features can be obtained.

[0063] (3) Predict the multi-level sub-output features separately to obtain multi-level sub-forecast results;

[0064] The prediction here actually refers to the upsampling process of the sub-output features. Since the prediction data dimensions of each sub-prediction result are different, for example, the output dimension of sub2 is (3, 128, 128), it needs to be restored to (3, 256, 256) through upsampling before it can be used to perform loss with the actual label (3, 256, 256) for model training.

[0065] The above multi-level sub-forecast results are as follows: Figure 2 sub1, sub2, sub3, sub4, sub5, sub6 shown in .

[0066] (4) The multi-level sub-forecast results are integrated to obtain the extrapolated forecast results.

[0067] By fusing sub1, sub2, sub3, sub4, sub5, and sub6, we can obtain the extrapolated forecast results.

[0068] See also Figure 5 The model training process and model prediction process diagram shown in the figure gives the model training and prediction process using data from 2021 to 2023. In the future, longer historical data can be accumulated to Figure 5 The process is used to perform operational modeling and forecasting. After the model is built and deployed, real-time forecasts can also be performed. First, the precipitation analysis field data (i.e., the live analysis data) is cropped from its original size (6, 1521, 1221) to (6, 256, 256) to enhance the model's ability to capture local precipitation details and facilitate subsequent model training. Model training is then performed, using the dual U-shaped structure U²-net model as a basis to design a new model structure (such as Figure 2 ), and then perform feature extraction and deep learning precipitation extrapolation on the cropped data. In order to fully integrate the advantages of different prediction methods, the extrapolation results obtained by the deep learning method are compared with the forecast results obtained by the traditional near-cross correlation extrapolation method ( Figure 5 The precipitation forecast field data) is fused through 1×1 convolution to output the final precipitation forecast result.

[0069] During the training process, the deep learning model parameters are continuously optimized through backpropagation, and a new deep learning precipitation near-extended intelligent forecasting model that can be stably applied to actual forecasting tasks is finally obtained. In the inference stage, only the precipitation data to be forecasted needs to be input. Through sequential cropping and sequential input, the model can automatically complete the adaptive weighted fusion of the deep learning extrapolation results and the traditional forecast results, thereby outputting a 1-3 hour high-precision precipitation sub-forecast result (3, 256, 256). Then, sequential splicing is performed to obtain the final 1-3 hour forecast result (3, 1521, 1221). The training process of the above intelligent forecasting model is as follows:

[0070] (1) Obtain a training sample set;

[0071] In practice, we can select several typical precipitation processes in complex regions and use the live and nowcast grid data provided by the China Meteorological Research Institute. Each sample data set includes an input image sequence consisting of 10-minute precipitation images for the past hour and forecast precipitation images for the next 1-3 hours, and an output image sequence consisting of hourly precipitation images for the next 3 hours.

[0072] For example, in a set of training samples, the input is a 10-minute precipitation image of the past hour, with a size of (6, 1521, 1221) + a fusion forecast precipitation image of the next 1-3 hours, with a size of (3, 1521, 1221), and the total input is (9, 1521, 1221); the output is a precipitation image of the next 3 hours (hourly), with a size of (3, 1521, 1221).

[0073] (2) The precipitation analysis field data and precipitation forecast field data in the training sample set are treated as a frame of image and cropped;

[0074] (3) Through the U²-net model, CA attention mechanism and hierarchical upsampling fusion module HUFM, feature extraction and near-field extrapolation processing are performed on the precipitation analysis field data after image cropping to obtain deep learning extrapolation forecast results; the deep learning extrapolation forecast results include multi-level sub-forecast results;

[0075] This process is similar to the steps in the aforementioned model prediction process. Please refer to the aforementioned content and will not be repeated here.

[0076] (4) Fusing the deep learning extrapolated forecast results with the precipitation forecast field data after image cropping to obtain the final forecast results;

[0077] (5) Based on the deep learning extrapolated forecast results, the final forecast results and the corresponding sample data, the loss value is calculated, and the deep learning model parameters are adjusted based on the loss value until the model converges to obtain a trained intelligent forecast model.

[0078] In this embodiment, the total loss of the model is calculated using MSE mean squared error loss (MSE) + MAE mean absolute error loss (MAE). In specific implementation, the loss value is calculated according to the following formula:

[0079]

[0080] in, 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; It 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-prediction result after sampling at the mth layer; Represents the weighted loss corresponding to the final forecast result; 、 They represent the weight corresponding to the m-th layer sampling and the weight of the final forecast result respectively; is the observed value of precipitation in the i-th sample, is the predicted value of precipitation in the i-th sample, n is the number of samples, is the weight of the i-th sample.

[0081] During the model training process provided in this example, Adam was used as the optimizer; the batch size was adjusted to 48; the total number of iterations was set to 80; the initial learning rate was set to 0.01, and a dynamic learning rate decay strategy was introduced. Specifically, if the performance metric on the validation set failed to improve three consecutive times during training, the learning rate would automatically decay to 50% of the original value. The model training platform was PyTorch. The model input was a continuous sequence of historical precipitation images, and the model output was hourly precipitation images for the next three consecutive hours.

[0082] Furthermore, this embodiment uses the commonly used evaluation metrics in the meteorological industry, namely TS (Threat Score), ETS (Equitable Threat Score), BS (Bias Score), and HSS (Heidke Skill Score), to objectively and quantitatively assess the accuracy of precipitation forecasts. TS represents the Threat Score, ETS represents the Equitable Threat Score, BS represents the Bias Score, and HSS represents the Heidke Skill Score. The Threat Score ranges from 0 (completely incorrect forecast) to 1 (perfect forecast), with higher values ​​indicating more accurate model forecasts. The Equitable Threat Score also ranges from 0 to 1, with higher values ​​indicating greater forecast skill and eliminating bias due to random chance. Bias Score values ​​closer to 1 indicate lower model bias, while values ​​greater than (less than) 1 indicate overly strong (weak) forecasts, wide (small) ranges, and a tendency towards false positives (missing predictions). The Heidke Skill Score ranges from -1 to 1, with 0 indicating a model's forecast skill comparable to a random forecast and 1 indicating perfect forecast skill. Negative values ​​indicate a model's skill worse than a random forecast. It is a comprehensive metric used to evaluate forecast model skill.

[0083] Under the condition of a precipitation threshold of 10 mm / h or above, the method and forecast model provided in this embodiment achieve a maximum improvement of 52.02% in TS and ETS scores, a maximum improvement of 22.22% in HSS scores, and a BS closer to the optimal value of 1 compared with the existing short-term precipitation forecast level. Compared with the traditional cross-correlation extrapolation method and the U-net and U²-net baseline deep learning models, the method and forecast model provided in this embodiment achieve significant performance improvements and extend the effective forecast time of precipitation impending extrapolation from 120 minutes to 180 minutes.

[0084] Unlike traditional deep learning precipitation extrapolation models, this embodiment generates preliminary extrapolation results and then dynamically fuses them with the 500-meter resolution numerical fusion forecast field using a 1×1 convolution. The fused precipitation forecast map is then used as feedback input for subsequent training and prediction processes, forming a closed-loop optimization mechanism. The advantages of this feedback mechanism are as follows:

[0085] (1) Enhance forecast accuracy and capture multi-scale information: The integration of numerical forecast results enables the model to not only capture heavy rainfall areas, but also obtain the simulation advantages of numerical models for large-scale precipitation systems, thereby improving the overall spatiotemporal consistency and forecast accuracy.

[0086] (2) Improve model robustness and reduce single model bias: Alleviate the problem of model error accumulation under complex weather conditions through information complementarity mechanism, and improve the adaptability to various precipitation types (convective, stratiform clouds).

[0087] (3) Extending the extrapolation time: The forecasting skill of the traditional extrapolation model decreases significantly after 60 minutes and almost completely loses its forecasting skill after 120 minutes. However, the fusion of numerical forecast results can effectively alleviate this problem and extend the effective forecast time from 120 minutes to 180 minutes.

[0088] The precipitation nowcasting intelligent forecasting method provided by the embodiment of the present application has the following significant advantages:

[0089] 1. Achieve 500-meter high-resolution deep learning precipitation nowcasting for 1–3 hours based on gridded meteorological data, which is a leading achievement in China;

[0090] 2. The HUFM module significantly improves image upsampling and enhances the prediction of localized heavy rainfall by introducing sub-pixel convolution and the SimAM attention mechanism.

[0091] 3. The CA mechanism strengthens the model's ability to model spatial contextual relationships and improves the accuracy of precipitation area positioning;

[0092] 4. By dynamically integrating numerical forecast information, the precipitation forecast time is effectively extended to 180 minutes, significantly improving the prediction capability and model stability for heavy precipitation events, and has good potential for commercial deployment.

[0093] Based on the above method embodiment, the present application embodiment also provides a precipitation near-extended intelligent forecasting device, see Figure 6 As shown, the device includes: a data acquisition module 62, which is used to obtain precipitation data to be forecasted; the precipitation data includes: actual data and forecast data; a clipping input module 64, which is used to clip the precipitation data and input the clipped precipitation data into the intelligent forecast model; the intelligent forecast model is a model obtained by training a specific model structure through a training sample set; wherein 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 in which a layered upsampling fusion module HUFM and a CA attention mechanism are added to the U²-net model; a forecast module 66 is used to perform feature extraction and deep learning precipitation extrapolation on the clipped actual data through the U²-net model, the CA attention mechanism and the layered upsampling fusion module HUFM in the intelligent forecast model to obtain an extrapolated forecast result, and adaptively weighted fuse the extrapolated forecast result with the forecast data to output a precipitation forecast result for the specified future period.

[0094] Furthermore, the above-mentioned prediction module 66 is used to perform a multi-level feature sampling operation on the actual data through the multi-level feature sampling U structure in the U²-net model to obtain multiple sampling features; in order from the lowest feature sampling U structure to the highest feature sampling U structure, the sampling features are sequentially fused through the CA attention mechanism and the layered upsampling fusion module HUFM, and the features are restored through the multi-level feature recovery U structure to obtain multi-level sub-output features; the multi-level sub-output features are predicted respectively to obtain multi-level sub-prediction results; and the multi-level sub-prediction results are fused to obtain an extrapolated prediction result.

[0095] Furthermore, the above-mentioned prediction module 66 is used to use the sampling features of the lowest-level feature sampling U structure as the current sub-output features of the current U structure, and perform the following feature fusion and recovery steps: process the current sub-output features through the layered upsampling fusion module HUFM to obtain the first feature; process the sampling features of the upper-level U structure of the current U structure through the CA attention mechanism module to obtain the second feature; fuse the first feature and the second feature to obtain the target feature; perform feature recovery processing on the target feature through the feature recovery U structure to obtain the sub-output features of the upper-level U structure; use the sub-output features of the upper-level U structure as the current sub-output features of the current U structure, and continue to perform feature fusion and recovery steps until multi-level sub-output features are obtained.

[0096] Furthermore, the above-mentioned layered upsampling fusion module HUFM includes: a first convolution layer, a sub-pixel convolution layer, a SimAM attention mechanism, a second convolution 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 feature through the first convolution layer; efficiently upsample the local features through the sub-pixel convolution layer to obtain upsampled features; perform feature processing on the upsampled features through the SimAM attention mechanism to obtain enhanced fusion features; feature processing includes: feature enhancement and feature fusion processing based on 3D weights to adaptively enhance key precipitation areas and edge details; perform a second convolution processing on the enhanced fusion features through the second convolution layer, and further optimize the feature expression of the features after the second convolution processing through the normalization layer and the ReLU activation layer to obtain the first feature.

[0097] Furthermore, the above-mentioned device also includes: a model training module, which is used to execute the following training process of the intelligent forecast model: obtaining a training sample set; treating the precipitation analysis field data and the precipitation forecast field data in the training sample set as a frame of image for cropping; performing feature extraction and proximity extrapolation processing on the precipitation analysis field data after image cropping through the U²-net model, CA attention mechanism and layered upsampling 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 the 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 forecast model.

[0098] Furthermore, the above model training module is used to calculate the loss value according to the following formula:

[0099]

[0100] in, 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; It 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-prediction result after sampling at the mth layer; Represents the weighted loss corresponding to the final forecast result; 、 They represent the weight corresponding to the m-th layer sampling and the weight of the final forecast result respectively; is the observed value of precipitation in the i-th sample, is the predicted value of precipitation in the i-th sample, n is the number of samples, is the weight of the i-th sample.

[0101] Furthermore, each of the sample data includes: an input image sequence consisting of 10-minute precipitation images for the past hour and predicted precipitation images for the next 1-3 hours, and an output image sequence including hourly precipitation images for the next 3 hours.

[0102] The device provided in the embodiment of the present application has the same implementation principle and technical effects as those in the aforementioned method embodiment. For the sake of brief description, for matters not mentioned in the embodiment of the device, reference can be made to the corresponding content in the aforementioned method embodiment.

[0103] The present application also provides an electronic device, such as Figure 7As shown, it is a structural diagram of the electronic device, wherein 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.

[0104] exist Figure 7 In the illustrated embodiment, the electronic device further includes a bus 72 and a communication interface 73 , wherein the processor 71 , the communication interface 73 and the memory 70 are connected via the bus 72 .

[0105] Among them, the memory 70 may include high-speed random access memory (RAM), and may also include non-volatile memory (non-volatile memory), such as at least one disk storage. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 73 (which can be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. can be used. The bus 72 can 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 ease of representation, Figure 7 Only one bidirectional arrow is used in the diagram, but this does not mean that there is only one bus or one type of bus.

[0106] The processor 71 can be an integrated circuit chip with processing capability. In implementation process, each step of the above method can be completed by integrated logic circuit of hardware in the processor 71 or by instructions in the form of software. The processor 71 described above can be a general processor, including a graphic processing unit (GPU), a central processing unit (CPU), a network processor (NP), etc.; can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, a discrete hardware component. The general processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as a hardware code processor to execute, or be executed by a combination of hardware and software modules in the code processor. The software module can be located in 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. The storage medium in the memory is read by the processor 71, and the hardware thereof is combined to complete the steps of the method of the foregoing embodiments.

[0107] The embodiment of the present application further provides a computer readable storage medium, which stores computer executable instructions. When the computer executable instructions are called and executed by a processor, the computer executable instructions cause the processor to implement the above method. For details, refer to the foregoing method embodiments, which will not be described here.

[0108] The computer program product of the method, the device and the electronic equipment provided by the embodiment of the present application includes a computer readable storage medium storing program codes. The instructions included in the program codes can be used to execute the method described in the foregoing method embodiments. For details, refer to the method embodiments, which will not be described here.

[0109] Unless otherwise specifically stated, the relative steps, numerical expressions and numerical values of the components and steps set forth in these embodiments do not limit the scope of the present application.

[0110] If the 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 that is executable by a processor. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0111] In the description of this application, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate the description of this application and simplify the description. They do not indicate or imply that the devices or components referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limitations on this application. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0112] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present application, which are used to illustrate the technical solutions of the present application, rather than to limit them. The scope of protection of the present application is not limited thereto. Although the present application has been described in detail with reference to the above-described embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily conceive of changes to the technical solutions described in the above-described embodiments within the technical scope disclosed in the present application, or perform equivalent replacements for some of the technical features thereof. These modifications, changes, or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application shall be subject to the scope of protection of the claims.

Claims

1. A precipitation nowcasting intelligent forecasting method, characterized in that: The method comprises: Obtaining precipitation data to be forecasted; the precipitation data includes: actual data and forecast data; The precipitation data is clipped and the clipped precipitation data is input into the intelligent forecast model; the intelligent forecast model is a model obtained by training a specific model structure through a training sample set; wherein the training sample set includes: historical precipitation data and corresponding actual precipitation data for a specified period in the future; the historical precipitation data includes: precipitation analysis field data and corresponding precipitation forecast field data; the specific model structure is Based on the model, the model structure of the layered upsampling fusion module HUFM and the CA attention mechanism is added; Through the intelligent forecasting model The model, CA attention mechanism and the layered upsampling fusion module HUFM perform feature extraction and deep learning precipitation extrapolation on the cropped real-time data to obtain an extrapolated forecast result, and adaptively weightedly fuse the extrapolated forecast result with the forecast data to output a precipitation forecast result within a specified future period; The layered upsampling fusion module HUFM includes: a first convolution layer, a sub-pixel convolution layer, a SimAM attention mechanism, a second convolution 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 layered upsampling fusion module HUFM to obtain the first feature includes: extracting the local features of the current sub-output feature through the first convolution layer; efficiently upsampling the local features through the sub-pixel convolution layer to obtain the upsampled feature; performing feature processing on the upsampled feature 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 areas and edge details; performing a second convolution processing on the enhanced fusion feature through the second convolution layer, and further optimizing the feature expression of the feature after the second convolution processing through the normalization layer and the ReLU activation layer to obtain the first feature.

2. The method according to claim 1, characterized in that Through the intelligent forecasting model The model, the CA attention mechanism, and the layered upsampling fusion module HUFM perform feature extraction and deep learning precipitation extrapolation on the cropped real-time data to obtain the extrapolated forecast result, including: Through the The multi-level feature sampling U structure in the model performs a multi-level feature sampling operation on the real-time data to obtain multiple sampling features; In order from the lowest feature sampling U structure to the highest feature sampling U structure, the sampled features are fused by the CA attention mechanism and the hierarchical upsampling fusion module HUFM, and the features are restored by the multi-level feature recovery U structure to obtain multi-level sub-output features; Predict the multi-level sub-output features separately to obtain multi-level sub-forecast results; The multi-level sub-forecast results are integrated to obtain the extrapolated forecast results.

3. The method according to claim 2, characterized in that The steps of fusing the sampled features in sequence from the lowest feature sampling U structure to the highest feature sampling U structure through the CA attention mechanism and the hierarchical upsampling fusion module HUFM, and recovering the features through the multi-level feature recovery U structure to obtain multi-level sub-output features include: The sampling features of the bottom-level feature sampling U structure are used as the current sub-output features of the current U structure, and the following feature fusion and recovery steps are performed: Processing the current sub-output feature through the hierarchical upsampling fusion module HUFM to obtain a first feature; Processing the sampling features of the U structure of the previous layer of the current U structure through the CA attention mechanism module to obtain the second feature; Fusing the first feature and the second feature to obtain a target feature; The target feature is subjected to feature recovery processing through a feature recovery U structure to obtain the sub-output feature of the upper-layer U structure; the sub-output feature of the upper-layer U structure is used as the current sub-output feature of the current U structure, and the feature fusion and recovery steps are continued until multi-level sub-output features are obtained.

4. The method according to claim 1, wherein The training process of the intelligent forecast model is as follows: Obtain a training sample set; The precipitation analysis field data and the precipitation forecast field data in the training sample set are respectively regarded as a frame of image and cropped; pass The model, CA attention mechanism and the hierarchical upsampling fusion module HUFM perform feature extraction and near-field 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; The deep learning extrapolated forecast result is integrated with the precipitation forecast field data after image cropping to obtain a final forecast result; Based on the deep learning extrapolated forecast results, the final forecast results and the corresponding sample data, a loss value is calculated, and the deep learning model parameters are adjusted based on the loss value until the model converges to obtain a trained intelligent forecast model.

5. The method according to claim 4, characterized in that The step of calculating the loss value based on the deep learning extrapolated prediction result, the final prediction result and the corresponding sample data includes: The loss value is calculated according to the following formula: in, 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-prediction result after sampling at the mth layer; Represents the weighted loss corresponding to the final forecast result; 、 They represent the weight corresponding to the m-th layer sampling and the weight of the final forecast result respectively; is the observed value of precipitation in the i-th sample, is the predicted value of precipitation in the i-th sample, n is the number of samples, is the weight of the i-th sample.

6. The method according to claim 4, characterized in that Each sample data includes: an input image sequence consisting of 10-minute precipitation images in the past hour and predicted precipitation images in the next 1-3 hours, and an output image sequence consisting of hourly precipitation images in the next 3 hours.

7. A precipitation near-fall extrapolation intelligent forecasting device, characterized in that: The device comprises: A data acquisition module is used to acquire precipitation data to be forecasted; the precipitation data includes: actual data and forecast data; The clipping input module is used to clip the precipitation data and input the clipped precipitation data into the intelligent forecast model; the intelligent forecast model is a model obtained by training a specific model structure through a training sample set; wherein the training sample set includes: historical precipitation data and the corresponding actual precipitation data of a specified period in the future; the historical precipitation data includes: precipitation analysis field data and the corresponding precipitation forecast field data; the specific model structure is Based on the model, the model structure of the layered upsampling fusion module HUFM and the CA attention mechanism is added; Forecast module, used for predicting the The model, CA attention mechanism and the layered upsampling fusion module HUFM perform feature extraction and deep learning precipitation extrapolation on the cropped real-time data to obtain an extrapolated forecast result, and adaptively weightedly fuse the extrapolated forecast result with the forecast data to output a precipitation forecast result within a specified future period; The layered upsampling fusion module HUFM includes: a first convolution layer, a sub-pixel convolution layer, a SimAM attention mechanism, a second convolution layer, a normalization layer, and a ReLU activation layer connected in sequence; the prediction module is also used to process the current sub-output feature of the current U structure through the layered upsampling fusion module HUFM to obtain a first feature, including: extracting local features of the current sub-output feature through the first convolution layer; efficiently upsampling the local features through the sub-pixel convolution 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 and feature fusion processing based on 3D weights to adaptively enhance key precipitation areas and edge details; performing a second convolution processing on the enhanced fusion feature through the second convolution layer, and further optimizing the feature expression of the feature after the second convolution processing through the normalization layer and the ReLU activation layer to obtain the first feature.

8. An electronic device, characterized in that: The method comprises a processor and a memory, wherein 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 6.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions prompt the processor to implement the method according to any one of claims 1 to 6.

Images