A high-precision minute-level precipitation nowcast method based on deep learning

Abstract

The application provides a high-precision minute-level precipitation nowcasting method based on deep learning, acquires various to-be-processed meteorological data; performs clipping processing on each kind of to-be-processed meteorological data, and outputs the clipped meteorological data; each kind of clipped meteorological data belonging to the same numbered position is combined, each group of obtained sub-meteorological data is input into a precipitation nowcasting model, and target precipitation prediction data in a second preset time period is output; in the mode, the precipitation nowcasting model can use independent encoders to encode different to-be-processed meteorological data separately, and then fuse the encoding results output by each encoder through a post-fusion module, so that the features of the various to-be-processed meteorological data can be complementary, the precipitation nowcasting precision can be improved, and then the demand of current meteorological services for regional sub-kilometer, minute-level fine-grained grid quantitative precipitation rapid prediction can be met.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and meteorology, and in particular to a high-precision minute-level nowcasting method for precipitation based on deep learning. Background Technology

[0002] Traditional short-term precipitation forecasts primarily rely on nowcasting (NW) and Numerical Weather Prediction (NWP) methods. However, each has its limitations. For instance, the accuracy of nowcasting decreases sharply with increasing forecast lead time. NWP forecast accuracy is highly dependent on initial and boundary conditions; even slight deviations in these conditions directly impact accuracy. Furthermore, NWP is typically calculated using supercomputers for Earth numerical simulations and is primarily used for long-term forecasts exceeding 24 hours, facing a "spin-up" problem in the last hour. Currently, even the most advanced NWP precipitation forecasts only achieve kilometer-level spatial resolution and hourly temporal resolution, limiting their application in fine-scale short-term precipitation forecasting and early warning. In addition, most existing AI-based precipitation forecasting models focus on hourly or seasonal precipitation evolution predictions. In summary, current technologies cannot meet the current meteorological operational needs for rapid quantitative precipitation forecasts at the sub-kilometer and minute levels. Summary of the Invention

[0003] The purpose of this invention is to provide a high-precision minute-level precipitation nowcasting method based on deep learning, so as to meet the current meteorological service requirements for rapid quantitative precipitation forecasting at the sub-kilometer and minute levels with fine gridded features.

[0004] This invention provides a high-precision minute-level precipitation nowcasting method based on deep learning. The method includes: acquiring multiple types of meteorological data to be processed; wherein, the multiple types of meteorological data to be processed include: precipitation analysis data and relative humidity analysis data within a first preset time period before the current time, and initial precipitation forecast data within a second preset time period after the current time; performing cropping processing on each type of meteorological data to be processed, and outputting cropped meteorological data corresponding to each type of meteorological data; combining each type of cropped meteorological data belonging to the same number position to obtain multiple sets of sub-meteorological data; and processing each set of sub-meteorological data... The data is input into a pre-trained precipitation nowcasting model to output target precipitation forecast data for a second preset time period. The precipitation nowcasting model includes a post-fusion module and multiple encoders corresponding to each type of meteorological data to be processed. Each encoder is used to encode the corresponding cropped meteorological data, and the encoded results are input into the post-fusion module for splicing and fusion processing. The data accuracy of the target precipitation forecast data is higher than that of the initial precipitation forecast data. The spatial resolution of the target precipitation forecast data is at the sub-kilometer level, and the temporal resolution is at the minute level.

[0005] Furthermore, the step of inputting each set of sub-meteorological data into a pre-trained precipitation nowcasting model, and outputting target precipitation forecast data for the second preset time period through the precipitation nowcasting model, includes: inputting each set of sub-meteorological data into a pre-trained precipitation nowcasting model, outputting precipitation forecast sub-data for the second preset time period corresponding to each set of sub-meteorological data through the precipitation nowcasting model; and splicing the precipitation forecast sub-data corresponding to all sub-meteorological data to obtain the target precipitation forecast data for the second preset time period.

[0006] Furthermore, each encoder includes multiple sequentially connected feature extraction layers. For each set of sub-meteorological data and each feature extraction layer in each encoder, the feature extraction layer performs double convolution processing on the received first data to obtain a first processing result. The first processing result is processed through the CBAM attention mechanism to output the feature extraction result corresponding to the feature extraction layer. The feature extraction result output by the last feature extraction layer is determined as the encoding result corresponding to the encoder and input to the post-fusion module. Wherein, if the feature extraction layer is the first feature extraction layer, the first data is the cropped meteorological data corresponding to the encoder in the set of sub-meteorological data; if the feature extraction layer is another feature extraction layer other than the first feature extraction layer, the first data is the feature extraction result output by the previous feature extraction layer. The post-fusion module is used to perform splicing and fusion processing on the encoding results output by each encoder to output the fused features corresponding to the set of sub-meteorological data.

[0007] Furthermore, the precipitation nowcasting model also includes a decoder; the decoder is used to decode the fusion features corresponding to each set of sub-meteorological data and output the precipitation forecast sub-data corresponding to each set of sub-meteorological data within the second preset time period.

[0008] Furthermore, the CBAM attention mechanism includes a channel attention submodule and a spatial attention submodule connected in sequence. The feature extraction layer is also used to: perform max pooling on the first processing result through the channel attention submodule to obtain a first pooling result; perform average pooling on the first processing result to obtain a second pooling result; perform fully connected processing on the first pooling result to obtain a max pooling output result; perform fully connected processing on the second pooling result to obtain an average pooling output result; sum the max pooling output result and the average pooling output result to obtain a summed result; process the summed result through a preset first activation function to obtain the importance coefficient of each channel in the first processing result, so as to base the importance coefficient on each channel. The importance coefficients of the channels are used to process the first processing result to obtain the second processing result. The second processing result is then subjected to max pooling through the spatial attention submodule to obtain the third pooling result. The second processing result is then subjected to average pooling to obtain the fourth pooling result. The third and fourth pooling results are concatenated to obtain the first concatenated result. The first concatenated result is then subjected to convolution to obtain the first convolution result. The first convolution result is then processed using a preset second activation function to obtain the importance coefficients of each spatial location in the second processing result. Based on the importance coefficients of each spatial location, the second processing result is processed to output the feature extraction result corresponding to that feature extraction layer.

[0009] Furthermore, the post-fusion module is also used to: concatenate the encoding results output by each encoder to output a second concatenation result; perform convolution processing on the second concatenation result to obtain a second convolution processing result; perform global average pooling processing on the second convolution processing result to obtain a fifth pooling result; perform fully connected processing on the fifth pooling result to obtain a fully connected processing result; process the fully connected processing result through a preset third activation function to output the attention weight of each channel in the second concatenation result; and process the second convolution processing result based on the attention weight of each channel to output the fusion feature corresponding to the set of sub-meteorological data.

[0010] Furthermore, the precipitation nowcasting model is trained as follows: Sample meteorological data is acquired; this sample meteorological data includes: precipitation analysis sample data and relative humidity analysis sample data for a first preset time period before a specified historical time, and initial precipitation forecast sample data and standard precipitation forecast data for a second preset time period after the specified historical time; the sample meteorological data is input into the initial model to output precipitation forecast prediction data for the second preset time period after the specified historical time; based on a preset target loss function, the loss value between the precipitation forecast prediction data and the standard precipitation forecast data is calculated; the initial model is trained based on the loss value until the loss value converges, resulting in a trained precipitation nowcasting model.

[0011] Furthermore, the objective loss function is: ; in, The mean squared error loss function represents the difference between the predicted precipitation data and the standard precipitation data. The weighted bias loss function represents the difference between precipitation forecast data and standard precipitation forecast data; This is the preset first weighting coefficient.

[0012] Furthermore, the weighted bias loss function is:

[0013] in, Indicates the first Precipitation classification threshold; Indicates the first Weighting coefficients corresponding to the precipitation classification thresholds; This indicates the number of precipitation classification thresholds; Indicates the order of the higher-order sigmoid function; This represents precipitation forecast data; This represents standard precipitation forecast data; u represents a step function. This is the preset second weighting coefficient.

[0014] This invention provides a high-precision minute-level precipitation nowcasting device based on deep learning. The device includes: an acquisition module for acquiring various meteorological data to be processed; wherein, the various meteorological data to be processed include: precipitation analysis data and relative humidity analysis data within a first preset time period before the current time, and initial precipitation forecast data within a second preset time period after the current time; a clipping module for clipping each type of meteorological data to be processed and outputting clipped meteorological data corresponding to each type of meteorological data; and a combination module for combining each type of clipped meteorological data belonging to the same number position to obtain multiple sets of sub-meteorological data. The output module is used to input each set of sub-meteorological data into a pre-trained precipitation nowcasting model, so that the precipitation nowcasting model can output target precipitation forecast data within a second preset time period. The precipitation nowcasting model includes a post-fusion module and multiple encoders corresponding to each type of meteorological data to be processed. Each encoder is used to encode the corresponding cropped meteorological data and input the encoded results into the post-fusion module for splicing and fusion processing. The data accuracy of the target precipitation forecast data is higher than that of the initial precipitation forecast data. The spatial resolution of the target precipitation forecast data is at the sub-kilometer level, and the temporal resolution is at the minute level.

[0015] This invention provides a high-precision minute-level precipitation nowcasting method based on deep learning, which acquires various types of meteorological data to be processed. These various types of meteorological data include: precipitation analysis data and relative humidity analysis data within a first preset time period before the current moment, and initial precipitation forecast data within a second preset time period after the current moment. Each type of meteorological data is cropped, and cropped meteorological data corresponding to each type of data is output. Cropped meteorological data belonging to the same position number are combined to obtain multiple sets of sub-meteorological data. Each set of sub-meteorological data is then input... The data is fed into a pre-trained precipitation nowcasting model to output target precipitation forecast data for a second preset time period. The precipitation nowcasting model includes a post-fusion module and multiple encoders corresponding to each type of meteorological data to be processed. Each encoder encodes the corresponding cropped meteorological data, and the encoded results are input into the post-fusion module for splicing and fusion. The accuracy of the target precipitation forecast data is higher than that of the initial precipitation forecast data. The spatial resolution of the target precipitation forecast data is sub-kilometer level, and the temporal resolution is minute level. In this approach, the precipitation nowcasting model can use independent encoders to encode different meteorological data to be processed separately, and then the encoded results output by each encoder are fused by the post-fusion module. This allows for feature complementarity among various meteorological data to be processed, improving the accuracy of precipitation nowcasting and thus meeting the current meteorological operational needs for rapid quantitative precipitation forecasting at the sub-kilometer and minute levels. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 A flowchart of a high-precision minute-level precipitation nowcasting method based on deep learning provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a precipitation nowcasting model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a CBAM attention mechanism provided in an embodiment of the present invention; Figure 4 A schematic diagram illustrating the training and forecasting structure of a precipitation nowcasting module provided in an embodiment of the present invention; Figure 5This is a schematic diagram of a high-precision minute-level precipitation nowcasting device based on deep learning, provided in an embodiment of the present invention. Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Traditional short-term precipitation forecasting primarily relies on nowcasting techniques (such as optical flow and cross-correlation methods) and NWP (Numeric Weather Prediction) methods. These methods are widely used in operational weather forecasting worldwide, but each has its limitations. Compared to traditional methods, deep learning-based artificial intelligence methods, with their nonlinear fitting capabilities, have been proven to better capture the spatiotemporal variations of meteorological elements using vast amounts of historical data, enabling predictions of weather system evolution. They have been widely used to improve traditional weather forecasts or directly for weather forecasting. However, most existing AI-based precipitation forecasting models focus on hourly or seasonal precipitation evolution predictions. Therefore, current technologies cannot meet the current operational needs of meteorological services for rapid, high-precision, gridded quantitative precipitation forecasts at the sub-kilometer and minute levels. Based on this, this invention provides a high-precision minute-level precipitation nowcasting method based on deep learning, which can be applied to scenarios requiring accurate forecasting of near-term precipitation events.

[0020] To facilitate understanding of this embodiment, a high-precision minute-level precipitation nowcasting method based on deep learning, disclosed in this embodiment of the invention, will first be introduced, such as... Figure 1 As shown, the method includes the following steps: Step S102: Obtain various meteorological data to be processed; wherein, the various meteorological data to be processed include: precipitation analysis data and relative humidity analysis data within a first preset time period before the current time, and initial precipitation forecast data within a second preset time period after the current time; The first and second preset time periods mentioned above can be set according to actual needs, and they can be the same or different, for example, both can be 1 hour; the precipitation analysis data mentioned above can be understood as the actual precipitation data that occurred within the first preset time period before the current time; the relative humidity analysis data mentioned above can be understood as the actual atmospheric humidity data within the first preset time period before the current time; the initial precipitation forecast data mentioned above can be understood as the future precipitation data within the second preset time period after the current time predicted by the traditional short-term precipitation forecast method; the data precision and accuracy of this initial precipitation forecast data are usually low.

[0021] In practical implementation, when it is necessary to forecast the imminent precipitation situation, the actual precipitation analysis data and actual relative humidity analysis data within a first preset time period before the current moment can be obtained, as well as the initial precipitation forecast data within a second preset time period after the current moment, forecasted using traditional short-term precipitation forecasting methods. For example, the new generation of rapid-refresh integrated seamless fusion forecasting system RISE (Rapid-refresh Integrated Seamless Ensemble), independently developed by the Meteorological Research Institute of City A, can provide gridded data generated in real time based on a large amount of observational data from the Beijing-Tianjin-Hebei region and mesoscale regional numerical model data. The types of analysis and forecasting data that the RISE system can provide include wind field, precipitation field, temperature and humidity field. In this embodiment, gridded precipitation data (corresponding to the above-mentioned precipitation analysis data), gridded relative humidity data (corresponding to the above-mentioned relative humidity analysis data), and initial precipitation forecast data can be obtained from the RISE system.

[0022] Step S104: Perform cropping on each type of meteorological data to be processed, and output the cropped meteorological data corresponding to each type of meteorological data to be processed. The original size of the meteorological data to be processed is usually large. For example, the size of gridded precipitation data obtained from the RISE system is usually 1521×1221. In order to reduce the complexity of data processing, each type of meteorological data to be processed is usually cropped. For example, it can be cropped to a size of 256×256 to obtain multiple cropped meteorological data corresponding to each type of meteorological data. Since the original size of each type of meteorological data to be processed is usually the same, the number of cropped meteorological data obtained after cropping according to the same cropping size is usually also the same.

[0023] Step S106: Combine each type of cropped meteorological data belonging to the same number position to obtain multiple sets of sub-meteorological data; For example, taking the same number position as the first row and first column, the cropped meteorological data of the first row and first column position after trimming the precipitation analysis data, the cropped meteorological data of the first row and first column position after trimming the relative humidity analysis data, and the cropped meteorological data of the first row and first column position after trimming the initial precipitation forecast data are combined to obtain the first set of sub-meteorological data; and so on, multiple sets of sub-meteorological data can be obtained.

[0024] Step S108: Input each set of sub-meteorological data into a pre-trained precipitation nowcasting model. The precipitation nowcasting model outputs target precipitation forecast data for a second preset time period. The precipitation nowcasting model includes a post-fusion module and multiple encoders corresponding to each type of meteorological data to be processed. Each encoder is used to encode the corresponding cropped meteorological data and input the encoded results into the post-fusion module for splicing and fusion processing. The data accuracy of the target precipitation forecast data is higher than that of the initial precipitation forecast data. The spatial resolution of the target precipitation forecast data is at the sub-kilometer level, and the temporal resolution is at the minute level.

[0025] The above-mentioned precipitation nowcasting model is based on the UNet model. To address the multi-source nature of the meteorological data to be processed, different meteorological data are encoded separately and then combined at the deepest semantic level of the model. In other words, this model is a deep learning model that constructs a multi-encoder fusion network structure. In this embodiment, the various meteorological data to be processed include three types, so the number of encoders is usually three. Each encoder is responsible for processing one type of meteorological data. For example, taking the first set of sub-meteorological data as an example, the cropped meteorological data corresponding to the first row and first column of precipitation analysis data can be input into the first encoder for encoding. The cropped meteorological data corresponding to the first row and first column of relative humidity analysis data can be input into the second encoder for encoding. The cropped meteorological data corresponding to the first row and first column of initial precipitation forecast data can be input into the third encoder for encoding. The three encoded results are then input into the post-fusion module for splicing and fusion. The same processing procedure is performed on each set of sub-meteorological data to obtain the target precipitation forecast data within the second preset time period. The data accuracy of the target precipitation forecast data is higher than that of the initial precipitation forecast data, and the accuracy will not decrease with the extension of the forecast lead time. Moreover, the spatial resolution of the target precipitation forecast data can reach the sub-kilometer level, and the temporal resolution can reach the minute level, for example, a spatial resolution of 500 meters and a temporal resolution of 10 minutes.

[0026] The aforementioned high-precision minute-level precipitation nowcasting method based on deep learning acquires various types of meteorological data to be processed. These include precipitation analysis data and relative humidity analysis data within a first preset time period before the current moment, and initial precipitation forecast data within a second preset time period after the current moment. Each type of meteorological data is cropped, outputting cropped meteorological data corresponding to each type. Cropped meteorological data belonging to the same position are combined to obtain multiple sets of sub-meteorological data. Each set of sub-meteorological data is input into a pre-trained precipitation nowcasting model to output target precipitation forecast data within the second preset time period. The precipitation nowcasting model includes a post-fusion module and multiple encoders corresponding to each type of meteorological data. Each encoder encodes the corresponding cropped meteorological data and inputs the encoded results into the post-fusion module for splicing and fusion. The precision of the target precipitation forecast data is higher than that of the initial precipitation forecast data. The spatial resolution of the target precipitation forecast data is sub-kilometer level, and the temporal resolution is minute level. In this approach, the precipitation nowcasting model can use independent encoders to encode different meteorological data to be processed separately, and then the encoding results output by each encoder are fused in the later stage through the post-fusion module. This can achieve feature complementarity of multiple meteorological data to be processed, improve the accuracy of precipitation nowcasting, and thus meet the current meteorological service needs for regional sub-kilometer and minute-level fine grid point quantitative precipitation rapid forecasting.

[0027] This invention also provides another high-precision minute-level precipitation nowcasting method based on deep learning. This method is implemented based on the method in the above embodiments and includes the following steps: Step 1: Acquire various types of meteorological data to be processed; among which, the various types of meteorological data to be processed include: precipitation analysis data and relative humidity analysis data within a first preset time period before the current time, and initial precipitation forecast data within a second preset time period after the current time; Step 2: Perform cropping on each type of meteorological data to be processed, and output the cropped meteorological data corresponding to each type of meteorological data. Step 3: Combine each type of cropped meteorological data belonging to the same number position to obtain multiple sets of sub-meteorological data; Step four: Input each set of sub-meteorological data into the pre-trained precipitation nowcasting model. The precipitation nowcasting model outputs the precipitation forecast sub-data corresponding to each set of sub-meteorological data within the second preset time period. The precipitation nowcasting model includes a post-fusion module and multiple encoders corresponding to each type of meteorological data to be processed. Each encoder is used to encode the corresponding cropped meteorological data and input the encoded results into the post-fusion module for splicing and fusion processing. Each encoder includes multiple sequentially connected feature extraction layers. For each set of sub-meteorological data and each feature extraction layer in each encoder, the feature extraction layer performs double convolution processing on the received first data to obtain a first processing result. The first processing result is processed through the CBAM (Convolutional Block Attention Module) attention mechanism, and the feature extraction result corresponding to the feature extraction layer is output. The feature extraction result output by the last feature extraction layer is determined as the encoding result corresponding to the encoder and input to the post-fusion module. Wherein, if the feature extraction layer is the first feature extraction layer, the first data is the cropped meteorological data corresponding to the encoder in the set of sub-meteorological data; if the feature extraction layer is another feature extraction layer other than the first feature extraction layer, the first data is the feature extraction result output by the previous feature extraction layer. like Figure 2 The diagram shows the structure of a precipitation nowcasting model. Each encoder includes four feature extraction layers, namely Layer 1 to Layer 4, as shown in the "Layer" module in the dashed box at the bottom of the figure. Each feature extraction layer performs a double convolution on the first received data to obtain the first processing result. Then, it performs a CBAM attention operation to simultaneously enhance spatial and channel features, generating the feature extraction results corresponding to each feature extraction layer (different color modules correspond to different encoders or feature types). For the first three feature extraction layers of each encoder, each feature extraction layer performs a downsampling operation, halving the feature map size and doubling the number of channels, before entering the next feature extraction layer. The feature extraction result output by the last feature extraction layer of each encoder is used as the encoding result of the corresponding encoder and input into the post-fusion module.

[0028] The post-fusion module is used to concatenate and fuse the encoding results output by each encoder, outputting the fused features corresponding to that set of sub-meteorological data. For each set of sub-meteorological data, the encoding results output by each encoder are concatenated and fused in the post-fusion module to form the fused features containing multi-source information corresponding to that set of sub-meteorological data.

[0029] The post-fusion module is also used for: splicing the encoding results output by each encoder to output a second splicing result; performing convolution processing on the second splicing result to obtain a second convolution processing result; performing global average pooling processing on the second convolution processing result to obtain a fifth pooling result; performing fully connected processing on the fifth pooling result to obtain a fully connected processing result; processing the fully connected processing result through a preset third activation function to output the attention weight of each channel in the second splicing result; and processing the second convolution processing result based on the attention weight of each channel to output the fusion feature corresponding to the set of sub-meteorological data.

[0030] like Figure 2 As shown in the dashed box at the bottom center, the Squeeze and Excite module, for each set of sub-meteorological data, after receiving the encoding results output by each encoder, can concatenate each encoding result to obtain a second concatenated result; after convolution processing (conv) on the second concatenated result, a second convolution processing result is obtained; then, through global average pooling, the spatial features of each channel are compressed (squeeze) into a one-dimensional vector (corresponding to the fifth pooling result mentioned above), and then this one-dimensional vector is input into a fully connected network for fully connected processing to obtain a fully connected processing result; after nonlinear activation (excitation) (corresponding to the processing of the fully connected processing result by the preset third activation function mentioned above), the attention weights of each channel in the second concatenated result are output; finally, the attention weights of each channel are multiplied (scaled) with the second convolution processing result channel by channel to obtain the fusion feature corresponding to the set of sub-meteorological data; the above channels can be understood as feature channels related to precipitation forecast, for example, one channel is specifically responsible for detecting "areas with large humidity gradients", and another channel is specifically responsible for detecting "specific data combinations that meet the requirements of a gigabit rainstorm", etc.

[0031] The precipitation nowcasting model also includes a decoder; the decoder is used to decode the fusion features corresponding to each set of sub-meteorological data and output the precipitation forecast sub-data corresponding to each set of sub-meteorological data in the second preset time period.

[0032] The post-fusion module is located between the feature output of multiple encoders and the input of the decoder. It is responsible for integrating the multi-source features extracted by different encoders and providing a unified "multi-source information fusion input" for the decoder. The fused features obtained by the post-fusion module are input into the decoder to provide multi-dimensional complementary information for the subsequent upsampling and feature reconstruction processes. After decoding by the decoder, a high-precision task result can be finally output, namely the precipitation forecast sub-data corresponding to each set of sub-meteorological data in the second preset time period. The specific processing of the decoder can be referred to relevant technologies, which will not be elaborated here.

[0033] CombinationFigure 2 As can be seen, the CBAM attention mechanism includes a channel attention submodule and a spatial attention submodule connected in sequence. In the overall process of CBAM, the first processing result after each layer performs a double convolution will first pass through the channel attention submodule to complete the channel dimension weighted optimization, and then be input into the spatial attention submodule for spatial dimension weighted adjustment, so as to achieve dual channel-space attention enhancement of features. Through the CBAM attention mechanism, the model's discriminative features and extraction capabilities for key channels and key spatial locations can be strengthened.

[0034] like Figure 3 The diagram illustrates the structure of a CBAM attention mechanism. The feature extraction layer is further configured to: perform max pooling on the first processing result through a channel attention submodule to obtain a first pooling result; perform average pooling on the first processing result to obtain a second pooling result; perform fully connected processing on the first pooling result to obtain a max pooling output result; perform fully connected processing on the second pooling result to obtain an average pooling output result; sum the max pooling output result and the average pooling output result to obtain a summed result; process the summed result using a preset first activation function to obtain the importance coefficient of each channel in the first processing result; and process the first processing result based on the importance coefficient of each channel to obtain a second processing result. like Figure 3 As shown, the feature extraction layer is also used to perform max pooling and average pooling on the first processing result, respectively. After inputting the two pooling results (i.e., the first pooling result and the second pooling result mentioned above) into the shared fully connected layer for processing, the max pooling output result and the average pooling output result are obtained respectively. The max pooling output result and the average pooling output result are added together and processed by the activation function (corresponding to the first activation function mentioned above) to obtain the importance coefficient of each channel in the first processing result. The importance coefficient of each channel is multiplied by the data of the corresponding channel in the first processing result to obtain the second processing result. This process can effectively capture the dependency relationship between the features of different channels and strengthen the feature representation of key channels related to precipitation.

[0035] The spatial attention submodule performs max pooling on the second processing result to obtain a third pooling result, and average pooling on the second processing result to obtain a fourth pooling result. The third and fourth pooling results are concatenated to obtain a first concatenated result. The first concatenated result is then convolved to obtain a first convolutional result. The first convolutional result is processed using a preset second activation function to obtain the importance coefficient of each spatial location in the second processing result. Based on the importance coefficient of each spatial location, the second processing result is processed to output the feature extraction result corresponding to this feature extraction layer.

[0036] like Figure 3 As shown, the feature extraction layer is also used to perform max pooling and average pooling on the second processing result output by the channel attention submodule, respectively. The two pooling results (i.e., the third pooling result and the fourth pooling result) are concatenated and then processed by the convolutional layer and the activation function (corresponding to the second activation function mentioned above) in turn to obtain the importance coefficient of each spatial location in the second processing result. The importance coefficient of each spatial location is multiplied by the data of the corresponding spatial location in the second processing result to obtain the feature extraction result corresponding to the feature extraction layer. This mechanism can accurately model important regions in spatial dimension and improve the model's ability to identify the spatial distribution and complex boundaries of precipitation targets.

[0037] The importance coefficients of the channels and spatial locations mentioned above refer to weighted averages across different dimensions (c, h, w), where c represents the channel, h represents the height, and w represents the width. The channel importance coefficient assigns different weights to different channels (c), while the spatial location importance coefficient assigns different weights to each point in the (h, w) space. Different importance coefficients represent different degrees of influence on the output. The CBAM convolutional attention mechanism simultaneously improves the channel and spatial representation capabilities of features without significantly increasing computational overhead, exhibiting excellent lightweightness and flexibility, making it suitable for accurate modeling of high-resolution precipitation features in complex scenarios.

[0038] The training process of the precipitation nowcasting module is described below. This precipitation nowcasting model can be trained through the following steps A to C: Step A: Obtain sample meteorological data; wherein, the sample meteorological data includes: precipitation analysis sample data and relative humidity analysis sample data within a first preset time period before the specified historical time, and initial precipitation forecast sample data and precipitation forecast standard data within a second preset time period after the specified historical time. The aforementioned precipitation forecast standard data refers to actual precipitation data within a second preset time period after a specified historical time; such as Figure 4 The diagram shows the training and forecasting structure of a precipitation nowcasting module; wherein, Figure 4The structure within the upper dashed box corresponds to the training process, while the structure within the lower dashed box corresponds to the actual forecasting process after the model is trained. In practice, 1521×1221 original grid sample data can be obtained from the RISE system. Considering that it is very difficult to directly use this size of original grid sample data for deep learning model training, the original grid sample data can be cropped to 256×256 to obtain the aforementioned meteorological sample data. For example, taking precipitation analysis data as an example, the original grid sample data can be cropped to 256×256 to obtain multiple 256×256 continuous sub-map sequences, i.e., the aforementioned precipitation analysis sample data. Among them, sequences with a precipitation value greater than 0.01 mm / 10 min accounting for more than 10% of the entire sequence can be selected for the training and validation sets. At the same time, the same method can be used to obtain the relative humidity analysis sample data, initial precipitation forecast sample data, and precipitation forecast standard data for the corresponding time. For example, Figure 4 As shown, taking a first preset time period and a second preset time period of 1 hour as an example, each set of meteorological data includes precipitation analysis sample data for 10 minutes every hour 1 hour before the specified historical time (e.g., Figure 4 The data includes "precipitation analysis data over the past 6 10-minute periods" and relative humidity analysis sample data (such as...). Figure 4 The data includes "relative humidity analysis data for the past 6 10-minute periods" and initial precipitation forecast sample data for one hour after a specified historical time (e.g., ...). Figure 4 The input data is “RISE precipitation forecast data for the next 6 1-10 minutes”, with a total input size of (18, 256, 256), where 18 represents a total of 18 sub-maps of 256×256.

[0039] Step B involves inputting sample meteorological data into the initial model to output precipitation forecast data for the second preset time period after a specified historical time. In practice, the sample meteorological data obtained above can be input into the initial model so that the precipitation forecast data for the second preset time period after a specified historical time can be predicted through the initial model.

[0040] Step C: Based on the preset target loss function, calculate the loss value between the precipitation forecast data and the standard precipitation forecast data. Train the initial model based on the loss value until the loss value converges to obtain the trained precipitation nowcasting model.

[0041] The predicted precipitation data and the standard precipitation forecast data (corresponding to the target loss function) can be calculated according to the preset target loss function. Figure 4The loss value is the difference between the predicted precipitation data and the standard precipitation forecast data ("the next 6 10-minute precipitation analysis data"). This loss value is used to measure the difference between the predicted precipitation data and the standard precipitation forecast data. The smaller the loss value, the smaller the difference between the predicted precipitation data and the standard precipitation forecast data, meaning that the initial model fits the sample meteorological data better. The initial model can be trained based on the loss value until the loss value converges, thus obtaining a trained precipitation nowcasting model. MSE Loss As shown in the structure of the lower dashed box, the processing procedure in the aforementioned embodiment can be referenced to use the trained precipitation nowcasting module to forecast the target precipitation forecast data.

[0042] For example, data from June to September 2022 and June to September 2023 can be used as the training and validation sets, respectively, while data from June to August 2024 can be used as the test set to test the final performance of the precipitation nowcasting model. This scheme adopts four precipitation forecast performance evaluation metrics, including: Threat Score (TS), Bias Score (BIAS), Probability of Detection (POD) score, and Equitable Threat Score (ETS). The meaning of each evaluation metric is explained below.

[0043] Accuracy TS Score: One of the most commonly used indicators for evaluating the accuracy of precipitation forecasts, also known as the Critical Success Index (CSI) score, it measures the consistency between predicted precipitation events and actual observed precipitation events. Its value ranges from 0 to 1, with higher values ​​indicating higher forecast accuracy.

[0044] Bias score (BIAS): Used to measure the forecast bias of precipitation events of a specific intensity, reflecting the overall effectiveness of precipitation forecasts at different levels. The closer the BIAS score is to 1, the better the overall forecast performance. The best forecast has a BIAS score of 1. A BIAS score greater than 1 indicates that there are too many false alarms in the precipitation nowcasting model, while a BIAS score less than 1 indicates that there are too many missed alarms.

[0045] The Point of Observation (POD) score measures the accuracy of precipitation forecasts in detecting actual precipitation events. It focuses on the proportion of correctly predicted events and is calculated as the ratio of the number of times the precipitation nowcasting model (a) hits the forecast to the total number of actual observed precipitation events of a certain magnitude or higher (a+c). The POD score ranges from 0 to 1. A higher POD score indicates a higher POD score, stronger detection capability for actual precipitation events, and better forecast performance.

[0046] Fair Threat (ETS) Score: This score takes into account the consistency between forecasts and observations due to random factors, thus providing a fairer evaluation of the performance of precipitation nowcasting models. A higher ETS score indicates more reliable precipitation nowcasting model forecasts.

[0047] Tests on the test set showed that, with a 10-minute lead time, the precipitation nowcasting model of this scheme achieved a threat TS score of 0.52-0.66 for precipitation thresholds of 0.1mm-5.8mm / 10min, which is 19%-120% higher than the traditional RISE extrapolation forecast. The hit rate POD score remained stable at 0.71-0.87, significantly better than the fluctuating and unstable forecast performance of directly using the UNet model. With a 60-minute lead time, the fair threat ETS score for medium and high thresholds (1.3mm-5.8mm / 10min) was 17%-22% higher than the traditional RISE extrapolation forecast, and the bias BIAS score was closer to the optimal value of 1, balancing the issues of missed and false alarms.

[0048] Furthermore, this scheme uses the target loss function LossFunction during the initial model training process, which is derived from the mean squared error loss function. WB Loss and weighted bias loss function Figure 5 The weighted calculation is obtained by comprehensively considering the overall data bias and precipitation score bias. The weight coefficients for the two are set as δ and (1-δ) respectively. The target loss function LossFunction is as follows: ; in, This represents the mean squared error loss function between the precipitation forecast data and the standard precipitation forecast data. The calculation formula for this mean squared error loss function can be found in relevant technologies. Based on this mean squared error loss function, the overall global error between the precipitation forecast data and the standard precipitation forecast data can be kept small, which helps the initial model training converge. The weighted bias loss function represents the difference between precipitation forecast data and standard precipitation forecast data; This is the preset first weighting coefficient, which can be set according to actual needs, such as 0.5.

[0049] The weighted bias loss function is as follows:

[0050] in, Indicates the first Precipitation classification threshold; Indicates the first Weighting coefficients corresponding to the precipitation classification thresholds; This indicates the number of precipitation classification thresholds; Indicates the order of the higher-order sigmoid function; This represents precipitation forecast data; This represents standard precipitation forecast data; u represents a step function. This is the preset second weighting coefficient.

[0051] In practical applications, to incorporate the bias score (BIAS) into the loss function, a higher-order sigmoid function can be used to approximate the step function, and then further used to obtain the skill score. The formula for the sigmoid function is shown below:

[0052] Where α represents the threshold for binary classification, β represents the order, and x represents the independent variable of the function, which in this scheme can be considered as the grid value of the precipitation grid map after being transformed by the model input. The sigmoid function formula is one of the common activation functions used to map the numerical values ​​of the model output to the range of 0-1.

[0053] The BIAS score, ranging from 0 to positive infinity, indicates a higher overall accuracy of the precipitation nowcasting model, with values ​​closer to 1. The loss function used for gradient descent typically shows a better model fit as smaller values, closer to 0, are considered better. Therefore, to improve the BIAS performance of precipitation forecasting by independently refining the loss function, this scheme designs the loss function value to be closer to 0 as the BIAS approaches 1. The loss function is shown in the following formula. As the BIAS moves away from 1, the loss value gradually increases; as the BIAS approaches 1, the loss value gradually approaches 0, indicating a good model fit.

[0054]

[0055] Furthermore, to reduce the potential errors arising from approximating the loss using a higher-order sigmoid function, this scheme also uses the original BIAS value to calculate the loss as supervision, denoted as Loss. hard Another differentiable loss value is denoted as Loss. soft The specific formula is shown below, where, This is the preset second weighting coefficient.

[0056] In addition, in precipitation forecasting tasks, the presence or absence of a forecast (determining whether it's sunny or rainy) and the forecast of heavy rain (highly hazardous) are two aspects of relatively high concern. Therefore, BiasLoss also considers the importance of precipitation at different thresholds, imposing greater penalties on deviations at the key 10-minute cumulative precipitation thresholds (0.1mm, 5.8mm), thereby helping to enhance the forecasting capabilities of AI precipitation models. For ease of understanding, the following formula lists the weighting coefficients corresponding to different thresholds:

[0057] The weighted bias loss function described above can be specifically described as follows: The Adam optimization algorithm was used for backpropagation during training optimization, with a total training iteration count of 100. Combined with batch size settings, the initial learning rate was set to 0.001, and a dynamic learning rate decay strategy was further introduced: if the performance metrics on the validation set fail to improve for three consecutive training iterations, the learning rate will automatically decay to 50% of its original value. This approach uses the deep learning framework PyTorch to construct the model structure.

[0058] Step 5: The precipitation forecast sub-data corresponding to all the sub-meteorological data are spliced ​​together to obtain the target precipitation forecast data within the second preset time period; wherein, the data accuracy of the target precipitation forecast data is higher than that of the initial precipitation forecast data; the spatial resolution of the target precipitation forecast data is at the sub-kilometer level and the temporal resolution is at the minute level.

[0059] After obtaining the precipitation forecast sub-data corresponding to each set of meteorological data within the second preset time period, all precipitation forecast sub-data can be spliced ​​together to obtain the target precipitation forecast data within the second preset time period after the current moment.

[0060] The proposed method utilizes high-resolution gridded data of the Beijing-Tianjin-Hebei region provided by the RISE system developed by the Meteorological Research Institute of City A. The data is uniformly segmented into (256, 256) sizes for training the precipitation nowcasting model. The model used in this method is a multi-encoder post-fusion network built on the UNet model. Precipitation analysis sample data, relative humidity analysis sample data, and initial precipitation forecast sample data from the RISE system are input into independent encoders for feature extraction. These are then combined using a "Squeeze and Excitation" module, and finally, spatial resolution is gradually restored in the decoder. Furthermore, in the training of the precipitation nowcasting model, the bias score (BIAS) from the skill score is incorporated into the loss function and participates in the model training process.

[0061] If data is fused at the input stage, it's called "early fusion," but this can lead to information entanglement when there are more complex dependencies between input variables. In contrast, this approach uses a network structure that addresses the multi-source nature of meteorological data by encoding different input data separately and performing "late fusion" at a deeper level of the model. This combination at the highest semantic level helps the model learn more complex dependencies between various pieces of information.

[0062] After training the precipitation nowcasting model using data of size (256, 256), when making actual forecasts, the original data of size (1521, 1221) can be sequentially divided into (256, 256), input into the trained precipitation nowcasting model to obtain extrapolation results, and then sequentially spliced ​​to obtain the cumulative precipitation data of size (1521, 1221) every 10 minutes for the last hour, which is the target precipitation forecast data mentioned above.

[0063] This solution aims to provide a deep learning-based method and model for nowcasting precipitation using high-resolution gridded meteorological data, specifically for forecasting cumulative precipitation every 10 minutes within 0-1 hour. This addresses the shortcomings of existing short-term nowcasting methods in terms of spatial resolution and temporal accuracy.

[0064] This scheme can be based on high-resolution gridded meteorological grid data with rapid updates at the 0.5 km and 10-minute intervals. Using the deep learning network UNet as the backbone, a multi-encoder "Late Fusion (LF)" deep learning model (corresponding to the above precipitation nowcasting model) can be designed to achieve nowcasting of cumulative precipitation in the Beijing-Tianjin-Hebei region from 0 to 1 hour every 10 minutes. This deep learning model constructs different encoders for different meteorological data, such as humidity. A "squeeze and excitation" module is introduced at a deeper level within the encoder to strengthen key features and execute a fusion strategy. A "CBAM (Convolutional Block Attention Module)" attention mechanism is introduced to enhance the network's ability to express important features through channel attention and spatial attention. For model training, the goal is to improve the accuracy of binary precipitation classification test indicators. A weighted bias loss function is designed by approximating step classification using a higher-order sigmoid function, and combined with the mean squared error loss function as the complete loss function used during model training. This improves the model's predictive ability for different minute-level precipitation thresholds, including light rain, moderate rain, heavy rain, and torrential rain. This approach can effectively improve the accuracy and stability of short-term precipitation nowcasting during the flood season in the Beijing-Tianjin-Hebei region, especially significantly improving the forecasting of moderate to heavy precipitation events.

[0065] The aforementioned high-precision minute-level precipitation nowcasting method based on deep learning has the following significant characteristics: 1. It can achieve deep learning-based 0-1 hour 10-minute precipitation forecasts based on high-precision gridded meteorological data with sub-kilometer spatial resolution and minute-level temporal resolution (such as 500-meter spatial resolution and 10-minute temporal resolution), which is at the forefront in China; it meets the current meteorological service needs for rapid quantitative precipitation forecasts with fine gridded data at the sub-kilometer and minute levels for regional areas.

[0066] 2. Constructing independent encoders for different meteorological data and performing late-stage fusion at a deeper level of the model helps the model learn more complex dependencies between various types of information.

[0067] 3. By approximating the step classification method using a higher-order sigmoid function, a weighted bias loss function is designed and combined with the mean squared error loss function as the complete loss function used during model training, thereby improving the model's predictive ability for different precipitation thresholds.

[0068] 4. The advantages of deep learning models, such as low training cost, short forecast start-up time, and good scalability, make them highly promising for operational deployment and they can be practically used in actual weather forecasting and early warning operations in the future.

[0069] This invention provides a high-precision minute-level precipitation nowcasting device based on deep learning, such as... Figure 6 As shown, the device includes: an acquisition module 50, used to acquire various meteorological data to be processed; wherein, the various meteorological data to be processed include: precipitation analysis data and relative humidity analysis data within a first preset time period before the current time, and initial precipitation forecast data within a second preset time period after the current time; a trimming module 51, used to trim each type of meteorological data to be processed, and output the trimmed meteorological data corresponding to each type of meteorological data; a combination module 52, used to combine each type of trimmed meteorological data belonging to the same number position to obtain multiple sets of sub-meteorological data; and an output module 53, used to output each... The meteorological data is input into a pre-trained precipitation nowcasting model to output target precipitation forecast data for a second preset time period. The precipitation nowcasting model includes a post-fusion module and multiple encoders corresponding to each type of meteorological data to be processed. Each encoder is used to encode the corresponding cropped meteorological data and input the encoded results into the post-fusion module for splicing and fusion. The data accuracy of the target precipitation forecast data is higher than that of the initial precipitation forecast data. The spatial resolution of the target precipitation forecast data is at the sub-kilometer level, and the temporal resolution is at the minute level.

[0070] In the aforementioned high-precision minute-level precipitation nowcasting device based on deep learning, the precipitation nowcasting model can encode different meteorological data to be processed separately using independent encoders. Then, the encoding results output by each encoder are fused in the later stage through a post-fusion module. This can achieve feature complementarity of multiple meteorological data to be processed, improve the accuracy of precipitation nowcasting, and thus meet the current meteorological service needs for rapid quantitative precipitation forecasting at the sub-kilometer and minute levels.

[0071] Furthermore, the output module 53 is also used to: input each set of sub-meteorological data into a pre-trained precipitation nowcasting model, and output precipitation forecast sub-data corresponding to each set of sub-meteorological data within a second preset time period through the precipitation nowcasting model; and splice together the precipitation forecast sub-data corresponding to all the sub-meteorological data to obtain the target precipitation forecast data within the second preset time period.

[0072] Furthermore, each encoder includes multiple sequentially connected feature extraction layers. For each set of sub-meteorological data and each feature extraction layer in each encoder, the feature extraction layer performs double convolution processing on the received first data to obtain a first processing result. The first processing result is processed through the CBAM attention mechanism to output the feature extraction result corresponding to the feature extraction layer. The feature extraction result output by the last feature extraction layer is determined as the encoding result corresponding to the encoder and input to the post-fusion module. Wherein, if the feature extraction layer is the first feature extraction layer, the first data is the cropped meteorological data corresponding to the encoder in the set of sub-meteorological data; if the feature extraction layer is another feature extraction layer other than the first feature extraction layer, the first data is the feature extraction result output by the previous feature extraction layer. The post-fusion module is used to perform splicing and fusion processing on the encoding results output by each encoder to output the fused features corresponding to the set of sub-meteorological data.

[0073] Furthermore, the precipitation nowcasting model also includes a decoder; the decoder is used to decode the fusion features corresponding to each set of sub-meteorological data and output the precipitation forecast sub-data corresponding to each set of sub-meteorological data within the second preset time period.

[0074] Furthermore, the CBAM attention mechanism includes a channel attention submodule and a spatial attention submodule connected in sequence. The feature extraction layer is also used to: perform max pooling on the first processing result through the channel attention submodule to obtain a first pooling result; perform average pooling on the first processing result to obtain a second pooling result; perform fully connected processing on the first pooling result to obtain a max pooling output result; perform fully connected processing on the second pooling result to obtain an average pooling output result; sum the max pooling output result and the average pooling output result to obtain a summed result; process the summed result through a preset first activation function to obtain the importance coefficient of each channel in the first processing result, so as to base the importance coefficient on each channel. The importance coefficients of the channels are used to process the first processing result to obtain the second processing result. The second processing result is then subjected to max pooling through the spatial attention submodule to obtain the third pooling result. The second processing result is then subjected to average pooling to obtain the fourth pooling result. The third and fourth pooling results are concatenated to obtain the first concatenated result. The first concatenated result is then subjected to convolution to obtain the first convolution result. The first convolution result is then processed using a preset second activation function to obtain the importance coefficients of each spatial location in the second processing result. Based on the importance coefficients of each spatial location, the second processing result is processed to output the feature extraction result corresponding to that feature extraction layer.

[0075] Furthermore, the post-fusion module is also used to: concatenate the encoding results output by each encoder to output a second concatenation result; perform convolution processing on the second concatenation result to obtain a second convolution processing result; perform global average pooling processing on the second convolution processing result to obtain a fifth pooling result; perform fully connected processing on the fifth pooling result to obtain a fully connected processing result; process the fully connected processing result through a preset third activation function to output the attention weight of each channel in the second concatenation result; and process the second convolution processing result based on the attention weight of each channel to output the fusion feature corresponding to the set of sub-meteorological data.

[0076] Furthermore, the precipitation nowcasting model is trained as follows: Sample meteorological data is acquired; this sample meteorological data includes: precipitation analysis sample data and relative humidity analysis sample data for a first preset time period before a specified historical time, and initial precipitation forecast sample data and standard precipitation forecast data for a second preset time period after the specified historical time; the sample meteorological data is input into the initial model to output precipitation forecast prediction data for the second preset time period after the specified historical time; based on a preset target loss function, the loss value between the precipitation forecast prediction data and the standard precipitation forecast data is calculated; the initial model is trained based on the loss value until the loss value converges, resulting in a trained precipitation nowcasting model.

[0077] Furthermore, the objective loss function is: ; in, The mean squared error loss function represents the difference between the predicted precipitation data and the standard precipitation data. The weighted bias loss function represents the difference between precipitation forecast data and standard precipitation forecast data; This is the preset first weighting coefficient.

[0078] Furthermore, the weighted bias loss function is:

[0079] in, Indicates the first Precipitation classification threshold; Indicates the first Weighting coefficients corresponding to the precipitation classification thresholds; This indicates the number of precipitation classification thresholds; Indicates the order of the higher-order sigmoid function; This represents precipitation forecast data; This represents standard precipitation forecast data; u represents a step function. This is the preset second weighting coefficient.

[0080] The high-precision minute-level precipitation nowcasting device based on deep learning provided in this embodiment of the invention has the same implementation principle and technical effects as the aforementioned high-precision minute-level precipitation nowcasting method embodiment based on deep learning. For the sake of brevity, any parts not mentioned in the embodiment of the high-precision minute-level precipitation nowcasting device based on deep learning can be referred to the corresponding content in the aforementioned high-precision minute-level precipitation nowcasting method embodiment based on deep learning.

[0081] This invention also provides an electronic device, see [link to relevant documentation]. Figure 6 As shown, the electronic device includes a processor 130 and a memory 131. The memory 131 stores machine-executable instructions that can be executed by the processor 130. The processor 130 executes the machine-executable instructions to implement the above-mentioned high-precision minute-level precipitation nowcasting method based on deep learning.

[0082] Furthermore, Figure 6 The electronic device shown also includes a bus 132 and a communication interface 133, with the processor 130, the communication interface 133 and the memory 131 connected via the bus 132.

[0083] The memory 131 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 133 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 132 may be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, ​ The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0084] Processor 130 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 130 or by instructions in software form. Processor 130 may be a general-purpose processor, including 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, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 131, and processor 130 reads the information in memory 131 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0085] This invention also provides a machine-readable storage medium storing machine-executable instructions. When these machine-executable instructions are called and executed by a processor, they cause the processor to implement the aforementioned high-precision minute-level precipitation nowcasting method based on deep learning. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0086] The computer program product of a high-precision minute-level precipitation nowcasting method based on deep learning provided in this embodiment of the invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here.

[0087] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A high-precision minute-level precipitation nowcasting method based on deep learning, characterized in that, The method includes: Acquire various types of meteorological data to be processed; wherein, the various types of meteorological data to be processed include: precipitation analysis data and relative humidity analysis data within a first preset time period before the current time, and initial precipitation forecast data within a second preset time period after the current time; Each type of meteorological data to be processed is cropped, and the cropped meteorological data corresponding to each type of meteorological data to be processed is output. The cropped meteorological data belonging to the same number position are combined to obtain multiple sets of sub-meteorological data; Each set of sub-meteorological data is input into a pre-trained precipitation nowcasting model to output target precipitation forecast data within the second preset time period. The precipitation nowcasting model includes a post-fusion module and multiple encoders corresponding to each type of meteorological data to be processed. Each encoder encodes the corresponding cropped meteorological data and inputs the encoded results into the post-fusion module for splicing and fusion. The precision of the target precipitation forecast data is higher than that of the initial precipitation forecast data. The spatial resolution of the target precipitation forecast data is at the sub-kilometer level, and the temporal resolution is at the minute level.

2. The method according to claim 1, characterized in that, The steps of inputting each set of sub-meteorological data into a pre-trained precipitation nowcasting model, and outputting target precipitation forecast data for the second preset time period through the precipitation nowcasting model, include: Each set of sub-meteorological data is input into a pre-trained precipitation nowcasting model, and the precipitation nowcasting model outputs the precipitation forecast sub-data corresponding to each set of sub-meteorological data within the second preset time period. The precipitation forecast sub-data corresponding to all the aforementioned sub-meteorological data are spliced ​​together to obtain the target precipitation forecast data within the second preset time period.

3. The method according to claim 1, characterized in that, Each encoder includes multiple feature extraction layers connected in sequence; For each group of sub-meteorological data and each feature extraction layer in each encoder, the feature extraction layer performs double convolution processing on the received first data to obtain a first processing result; the first processing result is processed through the CBAM attention mechanism to output the feature extraction result corresponding to the feature extraction layer; the feature extraction result output by the last feature extraction layer is determined as the encoding result corresponding to the encoder and input to the post-fusion module; wherein, if the feature extraction layer is the first feature extraction layer, the first data is the cropped meteorological data corresponding to the encoder in the group of sub-meteorological data; if the feature extraction layer is another feature extraction layer besides the first feature extraction layer, the first data is the feature extraction result output by the previous feature extraction layer; The post-fusion module is used to splice and fuse the encoding results output by each encoder, and output the fusion features corresponding to the group of sub-meteorological data.

4. The method according to claim 3, characterized in that, The precipitation nowcasting model also includes a decoder; The decoder is used to decode the fusion features corresponding to each group of sub-meteorological data and output the precipitation forecast sub-data corresponding to each group of sub-meteorological data within the second preset time period.

5. The method according to claim 3, characterized in that, The CBAM attention mechanism includes a channel attention submodule and a spatial attention submodule connected in sequence; the feature extraction layer is also used for: The channel attention submodule performs max pooling on the first processing result to obtain a first pooling result, and then performs average pooling on the first processing result to obtain a second pooling result. The first pooling result is processed by a fully connected layer to obtain the max pooling output result. The second pooling result is processed by a fully connected layer to obtain the average pooling output result; The max pooling output and the average pooling output are summed to obtain a summed result. The summation result is processed by a preset first activation function to obtain the importance coefficient of each channel in the first processing result. The first processing result is then processed based on the importance coefficient of each channel to obtain the second processing result. The spatial attention submodule performs max pooling on the second processing result to obtain a third pooling result, and then performs average pooling on the second processing result to obtain a fourth pooling result. The third pooling result and the fourth pooling result are concatenated to obtain the first concatenation result; The first splicing result is subjected to convolution processing to obtain the first convolution processing result; The first convolution processing result is processed by a preset second activation function to obtain the importance coefficient of each spatial location in the second processing result. The second processing result is then processed based on the importance coefficient of each spatial location to output the feature extraction result corresponding to the feature extraction layer.

6. The method according to claim 3, characterized in that, The post-fusion module is also used for: The encoding results output by each encoder are concatenated to output a second concatenated result; Perform convolution processing on the second splicing result to obtain the second convolution processing result; The second convolution result is subjected to global average pooling to obtain the fifth pooling result; The fifth pooling result is then subjected to full-connection processing to obtain the full-connection processing result; The fully connected processing result is processed by a preset third activation function, and the attention weight of each channel in the second stitching result is output. The second convolution processing result is then processed based on the attention weight of each channel, and the fusion feature corresponding to the group of sub-meteorological data is output.

7. The method according to claim 1, characterized in that, The precipitation nowcasting model was trained in the following way: Acquire sample meteorological data; wherein, the sample meteorological data includes: precipitation analysis sample data and relative humidity analysis sample data within a first preset time period before a specified historical time, and initial precipitation forecast sample data and precipitation forecast standard data within a second preset time period after the specified historical time. The sample meteorological data is input into the initial model so that the initial model can output precipitation forecast data for a second preset time period after the specified historical time. Based on a preset target loss function, the loss value between the precipitation forecast data and the precipitation forecast standard data is calculated. The initial model is trained based on the loss value until the loss value converges, thus obtaining a trained precipitation nowcasting model.

8. The method according to claim 7, characterized in that, The target loss function is: ； in, The mean squared error loss function represents the difference between the precipitation forecast data and the standard precipitation forecast data. The weighted deviation loss function represents the difference between the precipitation forecast data and the standard precipitation forecast data. This is the preset first weighting coefficient.

9. The method according to claim 8, characterized in that, The weighted bias loss function is: in, Indicates the first Precipitation classification threshold; Indicates the first Weighting coefficients corresponding to the precipitation classification thresholds; This indicates the number of precipitation classification thresholds; Indicates the order of the higher-order sigmoid function; This represents precipitation forecast data; This represents standard precipitation forecast data; u represents a step function. This is the preset second weighting coefficient.

10. A high-precision minute-level precipitation nowcasting device based on deep learning, characterized in that, The device includes: The acquisition module is used to acquire various meteorological data to be processed; wherein, the various meteorological data to be processed include: precipitation analysis data and relative humidity analysis data within a first preset time period before the current time, and initial precipitation forecast data within a second preset time period after the current time; The cropping module is used to crop each type of meteorological data to be processed and output the cropped meteorological data corresponding to each type of meteorological data to be processed. The combination module is used to combine each type of cropped meteorological data belonging to the same number position to obtain multiple sets of sub-meteorological data; The output module is used to input each set of sub-meteorological data into a pre-trained precipitation nowcasting model, so as to output target precipitation forecast data within the second preset time period through the precipitation nowcasting model; wherein, the precipitation nowcasting model includes a post-fusion module and multiple encoders corresponding to each type of meteorological data to be processed, each encoder is used to encode the corresponding cropped meteorological data, and input the obtained encoding result into the post-fusion module for splicing and fusion processing; the data accuracy of the target precipitation forecast data is higher than that of the initial precipitation forecast data; the spatial resolution of the target precipitation forecast data is at the sub-kilometer level, and the temporal resolution is at the minute level.

Images