Medium and short term rainstorm forecasting method based on deep learning and EC numerical forecasting

Abstract

The invention discloses a medium and short term rainstorm forecasting method based on deep learning and EC numerical forecasting, comprising the following steps: acquiring rainfall live observation data and EC numerical mode forecasting data over the years, and constructing an EC numerical mode forecasting sample data set based on the acquired data; forecasting model multi-channel data is obtained and preprocessed based on the EC numerical mode forecasting sample data set; and constructing an initial medium-short-term rainstorm forecasting model and a corresponding loss function, training the initial medium-short-term rainstorm forecasting model in combination with multi-channel data of the forecasting model, and obtaining a future precipitation lattice point forecasting field through the trained model. According to the method, the EC numerical forecasting data is used as the input of the Unet variant module, the rainstorm forecasting precision is improved by more than 15%, the weighted loss function is designed, the weight is given according to the rainfall, the high-precision forecasting of the rainstorm of more than 50mm is realized, and the accurate service is provided for disaster prevention and reduction.

Description

Technical Field

[0001] The present invention belongs to the technical field of weather forecasting, and in particular relates to a medium- to short-term rainstorm forecasting method based on deep learning and EC numerical forecasting. Background Art

[0002] Accurate rainstorm forecasts can provide early warnings of possible disasters such as floods, landslides, and mudslides, providing the government and relevant departments with valuable time to take disaster prevention and mitigation measures to reduce casualties and property losses. However, so far, the forecast information obtained by existing meteorological forecasters is not targeted enough. Among them, there is still a large gap between the TS score of the objective forecast of rainstorms in the short and medium term and the forecast demand for disaster prevention and mitigation of actual major meteorological disasters, which affects meteorological disaster prevention and mitigation decisions. The meteorological department urgently needs to improve the accuracy of weather forecasts.

[0003] At present, objective rainstorm forecasting has always been a difficult problem in operational forecasting, and its forecasting effect is relatively poor, especially the objective quantitative rainstorm forecasting effect is poor, mainly because rainstorm precipitation is not a continuous variable, nor is it an instantaneous observation value at a certain moment, but a cumulative value over a period of time. At the same time, precipitation is a variable with a serious skewed distribution. It is problematic to use traditional statistical methods to directly forecast precipitation. At present, scholars are constantly improving the numerical weather forecast model through various methods to improve the forecasting ability of rainstorms. However, compared with the prediction of atmospheric elements such as air pressure, temperature, humidity and wind, the prediction ability of numerical models for rainstorms is still far behind that of other atmospheric elements. Therefore, it is of great practical significance to study the correction forecasting technology of numerical model rainstorm prediction. Practice has also proved that by correcting the original rainstorm precipitation products of numerical models, more accurate rainstorm forecasts can usually be provided. So far, a large number of technical studies on the use of numerical forecast products for objective rainstorm forecast correction have been carried out both at home and abroad. However, these revised forecast methods often cannot well characterize the complex nonlinear dynamic processes involved in the rainstorm system when surveying forecast factors and constructing the mapping relationship between forecast factors and forecast objects, nor can they make good use of the various forecast feature information provided by the model. Therefore, they have great limitations in practical applications. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a medium- and short-term rainstorm forecasting method based on deep learning and EC numerical forecasting to solve the problems existing in the above-mentioned prior art.

[0005] To achieve the above object, the present invention provides a medium- and short-term rainstorm forecasting method based on deep learning and EC numerical forecasting, comprising:

[0006] Obtain the actual rainfall observation data and EC numerical model forecast data of previous years; construct the EC numerical model forecast sample data set based on the actual rainfall observation data and EC numerical model forecast data of previous years;

[0007] Obtaining forecast model multi-channel data based on the EC numerical model forecast sample data set, and preprocessing the forecast model multi-channel data;

[0008] An initial short- to medium-term rainstorm forecast model is constructed based on the Unet structure, and a model training loss function is constructed. The initial short- to medium-term rainstorm forecast model is trained based on the model training loss function and the preprocessed forecast model multi-channel data to obtain the final short- to medium-term rainstorm forecast model; based on the final short- to medium-term rainstorm forecast model, a grid forecast field of future precipitation is obtained.

[0009] Optionally, the actual rainfall observation data is the 24-hour accumulated precipitation data of the automatic weather station; the EC numerical model forecast data includes rainfall forecast fields of several ensemble members of the EC ensemble forecast and U and V wind field data of multiple levels of EC fine grids.

[0010] Optionally, the process of constructing the EC numerical model forecast sample dataset includes:

[0011] Each starting time within the time range corresponding to the actual rainfall observation data is taken as the starting time of precipitation in the next 24 hours, and the EC numerical model forecast data before the starting time is extracted; the EC numerical model forecast data is screened according to the precipitation forecast requirements, and after screening, it is combined with the precipitation to obtain the EC numerical model forecast sample data set.

[0012] Optionally, the forecast model multi-channel data includes rainfall forecast field input channel data and wind field forecast field input channel data.

[0013] Optionally, the process of obtaining multi-channel data of the forecast model includes:

[0014] The rainfall forecast fields of several ensemble members of the EC ensemble forecast are reconstructed and K member rainfall forecast fields are selected. The reconstruction process includes: sorting the forecast values ​​of several ensemble members corresponding to each grid point from large to small, selecting the first K members respectively, and obtaining the K member rainfall forecast fields corresponding to each grid point as the input channel data of the rainfall forecast field; based on the U and V wind field data at three levels of 500hPa, 850hPa and 925hPa, the wind speed and wind direction of the 12h and 36h forecast fields at each level are calculated respectively to obtain the input channel data of the wind field forecast field.

[0015] Optionally, the latitude and longitude range of the model input is determined based on the demand for forecast information, and the multi-channel data of the forecast model is scaled according to the latitude and longitude range.

[0016] Optionally, based on the Unet structure, an attention mechanism and a dilated convolution module are introduced to obtain a Unet variant as an initial short- to medium-term rainstorm forecast model.

[0017] Optionally, the model training loss function is:

[0018]

[0019] Where loss represents the total loss; N is the total number of samples; w s is the weight of the sth sample, usually used to adjust the loss of the minority class; y s The true label of the sth sample; The predicted label for the sth sample.

[0020] Compared with the prior art, the present invention has the following advantages and technical effects:

[0021] The present invention first obtains the actual observation data of rainfall in previous years and the EC numerical model forecast data; constructs the EC numerical model forecast sample data set based on the actual observation data of rainfall in previous years and the EC numerical model forecast data; then, obtains the forecast model multi-channel data based on the EC numerical model forecast sample data set, and pre-processes the forecast model multi-channel data; then, constructs an initial medium- and short-term heavy rain forecast model and a corresponding loss function, trains the initial medium- and short-term heavy rain forecast model based on the model training loss function and the pre-processed forecast model multi-channel data, and obtains the final medium- and short-term heavy rain forecast model; and obtains the future precipitation grid forecast field based on the final medium- and short-term heavy rain forecast model.

[0022] The present invention uses EC numerical model forecast data before the start time as EC ensemble forecast member rainfall forecast fields, EC fine grid 500hPa, 850hPa and 925hPa U, V wind field data, and obtains a preliminary numerical model forecast product data set by processing and calculating the time and space resolution of each element consistent with the precipitation forecast requirements. The obtained data set is used to construct multi-channel input data of the model, which is used as the input of the Unet variant module, and finally the output of the entire generator is obtained. The forecasting ability of the forecast model for heavy rains above 50mm is improved by more than 15% compared with the current heavy rain forecast level, which can effectively improve the forecast accuracy of medium- and short-term heavy rains, has good application prospects, and provides more accurate forecast services for disaster prevention and mitigation.

[0023] In view of the uneven distribution of forecast objects, the present invention designs a weighted loss function, which is the sum of the weighted losses of the weighted mean absolute error and the weighted mean square error. The weighting coefficient here is to assign different weights according to the actual rainfall value of the site, that is, the larger the rainfall value, the greater the weight assigned. The information feedback capacity of the rainstorm area is increased through the above-mentioned weight assignment rule so that it will not be averaged out by a large area of ​​weak precipitation area, and the high-precision forecast of the heavy rain and heavy precipitation echo above 50mm is achieved, which has higher forecast accuracy in the early warning of disaster prevention and mitigation. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0025] Figure 1 A flow chart of a method for forecasting short- to medium-term rainstorms according to an embodiment of the present invention;

[0026] Figure 2 A schematic diagram of precipitation forecast according to an embodiment of the present invention;

[0027] Figure 3 Schematic diagram of actual precipitation according to an embodiment of the present invention. DETAILED DESCRIPTION

[0028] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0029] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0030] Embodiment 1

[0031] like Figure 1-3 As shown, this embodiment provides a medium- and short-term rainstorm forecasting method based on deep learning and EC numerical forecasting, including:

[0032] Obtain the actual rainfall observation data and EC numerical model forecast data of previous years; construct the EC numerical model forecast sample data set based on the actual rainfall observation data and EC numerical model forecast data of previous years;

[0033] Specifically, the actual rainfall observation data of ground meteorological stations are obtained from the TianQing database, which is the actual rainfall for each station every 24 hours over the years; the EC numerical model forecast data are the rainfall forecast field data of 51 ensemble members of the EC ensemble forecast obtained from the MICAPS distributed system issued by the China Meteorological Administration, as well as the U and V wind field data of the EC fine grid 500hPa, 850hPa and 925hPa levels.

[0034] Furthermore, the process of constructing the EC numerical model forecast sample data set includes:

[0035] Each starting time within the time range corresponding to the historical actual rainfall data is taken as the starting time of precipitation in the next 24 hours, and the forecast data of the numerical model starting 12 hours before the starting time is obtained: the rainfall forecast fields of 51 ensemble members of the EC ensemble forecast and the U and V wind field data of 500hPa, 850hPa and 925hPa of the EC fine grid. The rainfall forecast field of the EC ensemble forecast and the U and V wind fields of 500hPa, 850hPa and 925hPa of the EC fine grid are interpolated and calculated at a resolution of 5 km within the forecast range, and the samples with all sunny days within the forecast range are eliminated. After completing all the screening, the numerical model forecast sample data set containing rainfall forecast fields and wind fields is finally obtained.

[0036] Exemplarily, the data used are the rainfall forecast field data of 51 ensemble members of the EC ensemble forecast and the U and V wind field data of the EC fine grid 500hPa, 850hPa and 925hPa levels, and the sampling period is May 2020-June 2024. Based on the fact that the numerical model has been continuously improved and optimized in recent years, in order to enable the model to obtain the latest model data as possible for training during the training process, in terms of the division of the data set, May-July 22 is used as the test data set, May 20-April 22, August 22-December 23 and January-April 24 are used as training sample data sets, and April-August 24 is used as a completely independent sample for testing.

[0037] Obtaining forecast model multi-channel data based on the EC numerical model forecast sample data set, and preprocessing the forecast model multi-channel data;

[0038] In some embodiments, the forecast model multi-channel data includes rainfall forecast field input channel data and wind field forecast field input channel data.

[0039] In some embodiments, the process of obtaining multi-channel data of the forecast model includes:

[0040] The rainfall forecast fields of the 51 ensemble members of the EC ensemble forecast are reconstructed and K member rainfall forecast fields are selected. The reconstruction process includes: sorting the forecast values ​​of the 51 ensemble members at each grid point from large to small, selecting the first K members respectively, and obtaining the K member rainfall forecast fields corresponding to each grid point as the input channel data of the rainfall forecast field; based on the U and V wind field data at the three levels of 500hPa, 850hPa and 925hPa, the wind speed and wind direction of the 12h and 36h forecast fields at each level are calculated respectively to obtain the input channel data of the wind field forecast field.

[0041] In some embodiments, the latitude and longitude range of the model input is determined based on the demand for forecast information, and the multi-channel data of the forecast model is scaled according to the longitude and latitude range.

[0042] Specifically, (1) Rainfall forecast channel data: The EC ensemble forecast rainfall forecast field has 51 members. Since we are going to build a medium- to short-term rainstorm forecast model, and in order to reduce the resource consumption during the training of the forecast model, we first reconstruct all 51 forecast field members of the ensemble forecast model, that is, sort the forecast values ​​of the 51 members of each grid point from large to small, select the first K members according to the sorting, and obtain a new K member forecast field (the members selected for each grid point (the original 51 members) are not necessarily the same). Finally, the forecast fields of the newly obtained K members are used as part of the model input channel (K rainfall forecast field channels). (2) Wind field data channel data: U and V forecast field data of the European Center fine grid 500hPa, 850hPa and 925hPa altitude layers, taking U\V at 500hPa as an example (the same applies to 850hPa and 925hPa). In view of the lag problem of EC data in actual business, when actually forecasting precipitation in the next 24 hours, the forecast field data of the previous 12-36 hours is used for forecasting. Therefore, when calculating the U\V wind field data, the U wind and V wind are used to calculate the wind speed and wind direction (0-360°) of the 12h and 36h forecast fields of the height level respectively, that is, 2 wind directions and 2 wind speed channel data. Similarly, after calculating the U\V at the 850hPa and 925hPa height layers, 2 wind directions and 2 wind speed channel data can be obtained respectively. In summary, a total of 6 wind direction and 6 wind speed channel data, a total of 12 wind field channel data as the input of the model.

[0043] Specifically, the newly obtained K rainfall forecast fields are used as part of the model input channel, and the 12 wind field data are used as other model channel data. A total of K+12 channel data are used as the input of the Unet variant module to finally obtain the output of the model.

[0044] An initial short- to medium-term rainstorm forecast model is constructed based on the Unet structure, and a model training loss function is constructed. The initial short- to medium-term rainstorm forecast model is trained based on the model training loss function and the preprocessed forecast model multi-channel data to obtain the final short- to medium-term rainstorm forecast model; based on the final short- to medium-term rainstorm forecast model, a grid forecast field of future precipitation is obtained.

[0045] In some embodiments, based on the Unet structure, an attention mechanism and a dilated convolution module are introduced to obtain a Unet variant as an initial medium- and short-term heavy rain forecast model.

[0046] Specifically, the main structure of the model is generally an adversarial network structure, and the generator is mainly a Unet variant module. This embodiment adopts an attention mechanism module and an ASPP module based on a simple Unet network to form a variant module. The specific operation process of the model is as follows: the newly obtained K rainfall forecast fields are used as part of the model input channel, and the 12 wind field data are used as other model channel data. A total of K+12 channel data are used together as the input of the Unet variant module, and finally the output of the model is obtained. When the model is trained, this Unet variant module uses a weighted loss function to update the model.

[0047] Specifically, in order to better train the model, this embodiment transforms the loss function. In the training of the model, the role of the loss function is to evaluate the fit of the current training of the model and to modify its weight according to the value of the loss function.

[0048] A weighted loss function is designed for the generator, which is the sum of the weighted losses of the weighted mean absolute error and the weighted mean square error. Its general form can be expressed as:

[0049]

[0050] Where loss represents the total loss; N is the total number of samples; w s is the weight of the sth sample, usually used to adjust the loss of the minority class; y s The true label of the sth sample; The predicted label for the sth sample.

[0051] The weighting coefficient is assigned different weights according to the actual rainfall. That is, the greater the rainfall, the greater the weight. The purpose of this is to increase the information feedback ability of the heavy rainfall area so that it is not averaged by a large area of ​​weak rainfall area. Set the actual rainfall size to 6 levels:

[0052] Level 1 rainfall: 1-5mm, level 2 rainfall: 5-10mm, level 3 rainfall: 10-40mm, level 4 rainfall: 40-50mm, level 5 rainfall: 50-100mm, level 6 rainfall: ≥100mm.

[0053] Step S7, training and obtaining a prediction model;

[0054] Step S8, calculate and obtain the grid point forecast of future heavy rain precipitation, and generate a precipitation grid point forecast field.

[0055] This embodiment mainly constructs an adversarial network deep learning model based on EC multi-forecast feature channel input and Unet variant module; in the selection of model input, the rainfall forecast field channel data of EC ensemble forecast and the U, V forecast field channel data of EC fine grid 500hPa, 850hPa and 925hPa height layers are selected; in the model input preprocessing, the input data is scaled; the main network structure is based on Unet, and the attention mechanism and the void convolution module are added to improve the feature extraction and extrapolation capabilities of the model; in order to better train this model, the loss function is further adaptively modified.

[0056] The following example uses the precipitation forecast for the next 24 hours starting at 08:00 on September 7, 2024 as an example:

[0057] (1) Collect sample data. The statistical period is from May 2020 to June 2024, of which the samples from May 2020 to April 2022, August 2022 to December 2023, and January to April 2024 are used as training samples, May to July 2022 is used as a test set, and April to September 2024 is tested as a completely independent sample. Consistent with the business reporting time, this embodiment sets 08:00 and 20:00 Beijing time as the reporting time, and the precipitation in the next 24 days after the reporting time is the forecast object. The collected data are: the actual observation data of ground rainfall at the ground meteorological station every 24 hours, the U and V forecast field data of the 51 rainfall forecast fields of the EC ensemble forecast and the European Center fine grid 500hPa, 850hPa and 925hPa altitude layers. Among them, the actual observation data of ground rainfall is the precipitation in the past 24 hours.

[0058] (2) Model input preprocessing. The test object of this embodiment is the rainstorm forecast situation with a grid spacing of 5 kilometers in the area of ​​20-27°N, 104.4-112.5°E. Considering that the rainstorm process in this area is often formed by the frontal rainstorm formed by the convergence of the southern tropical monsoon system and the northern cold air moving south, these are system processes with a relatively large range. Therefore, in the process of constructing the model input, the strategy of adopting the latitude and longitude range of the model input elements is slightly larger than the scope of the study area. In the model input of this embodiment, the range of the rainfall forecast field and the wind field are both intercepted from the data within the range of 19.5-28°N, 104-112.5°E of the model, and further interpolated into 0.05*0.05 grid data, and the obtained model input matrix is ​​(batchsize,c,171,171), where batchsize is the batch size of the training, c is the number of channels, and the height and width of the grid data are both 171.

[0059] (4) Construct the input of the forecast model. Sort the forecast values ​​of the 51 EC ensemble forecast model members at each grid point from large to small to obtain new K (K = 10) rainfall forecast field channels s1_data; 12 wind field channel data s2_data of 6 wind directions and 6 wind speeds at the European Central Fine Grid 500hPa, 850hPa and 925hPa altitude layers; a total of K+12=22 channel data are used to construct the final forecast model input matrix s3_data. At this time, the dimension of the s3_data tensor is (batchsize, 22, 171, 171), where batchsize is the batch size of the training input, 22 is 22 channel data, and 171 is the width and height of the data. Since the data size of 171 is not very friendly to convolution (it is easy to cause the height and width of the grid data to be unequal), the input data of the model is scaled before model training in this example, that is, (batchsize, 22, 171, 171) is scaled to (batchsize, 22, 256, 256). After the model training is completed, the obtained model output is (batchsize, 1, 256, 256) and scaled back to the original (batchsize, 1, 171, 171).

[0060] (5) Construct the EAAUnet medium- and short-term heavy rain forecast model. The K+12 channel input matrix s3_data is used as the input of the EAAUnet model, and it is further trained and the optimal model is saved. Among them, the weighted loss function is used for training during model training: that is, according to the rainfall size classification: first-level rainfall: 1-5mm, second-level rainfall: 5-10mm, third-level rainfall: 10-40mm, fourth-level rainfall: 40-50mm, fifth-level rainfall: 50-100mm, sixth-level rainfall: ≥100mm. The weights assigned to each level are: first-level rainfall: 1, second-level rainfall: 2, third-level rainfall: 5, fourth-level rainfall: 8, fifth-level rainfall: 10, sixth-level rainfall: 20;

[0061] The final loss function is:

[0062]

[0063] (6) Calculate the latest forecast sample. Taking 08:00 on September 7, 2024 as the starting time, when forecasting the precipitation forecast for the next 24 hours, the model forecast data used is the EC ensemble forecast starting at 20:00 on September 6, 2024 and the 12-36h forecast field of the EC fine grid model. The specific calculation of each channel data is as follows:

[0064] 1) Rainfall data channel: The accumulated rainfall forecast field of the EC ensemble forecast model is downloaded in real time from the micaps distributed database. The accumulated rainfall forecast field of 36h (the next 36 hours at 20:00 on September 6, 2024) is subtracted from the accumulated rainfall forecast field of 12h (the next 12 hours at 20:00 on September 6, 2024) to obtain the accumulated rainfall forecast field for the forecast period (the next 24 hours at 08:00 on September 7, 2024), which is recorded as R0. First, R0 is sliced ​​according to the research range of 19.5-28°N, 104-112.5°E, and the data extracted after slicing are interpolated into a data matrix with a grid spacing of 0.05*0.05, denoted as R1. At this time, the data matrix dimension of R1 is (51, 171, 171), where 51 is 51 members and 171 is the number of grid points in the latitudinal and radial directions, the same below; secondly, the 51 forecast values ​​of each grid point of R1 are sorted, and the maximum 10 forecast values ​​of each grid point are extracted to construct a new forecast field of 10 forecast members, denoted as R2. At this time, the data matrix dimension of R2 is (10, 171, 171).

[0065] 2) Wind channel data: The U and V wind forecast fields of the EC fine grid 500hPa, 850hPa and 925hPa altitude layers will be downloaded in real time from the micaps distributed database. First, the wind speed and wind direction (0-360°) of the 12h and 36h forecast fields of each altitude layer are calculated respectively, so as to obtain 12 wind channel data of 6 wind speeds and 6 wind directions, which are denoted as W0; secondly, W0 is sliced ​​according to the research range of 19.5-28°N, 104-112.5°E, and the data extracted after slicing are interpolated into a data matrix of 0.05*0.05 grid spacing, which is denoted as W1. At this time, the data matrix dimension of W1 is (12, 171, 171)

[0066] 3) Final input data of the model: Concatenate R2 and W1 by channel to construct a final model input X. At this time, the data matrix dimension of X is (22, 171, 171).

[0067] (7) Forecast using the trained model. First, the model input X obtained in (6) is scaled to the shape of (1, 22, 256, 256) (where 1 is the size of the batch), and then input into the trained model to obtain a rainfall forecast field R', whose dimension is (1, c, 256, 256) (where c = 1, is the number of channels, the same below); secondly, R' is scaled back to (1, c, 171, 171), which is R ’ The last revised rainfall forecast field.

[0068] (8) Calculation of actual grid point precipitation of the model. Calculate the precipitation of the grid point in the next 24 hours and complete the forecast of precipitation of the grid point in the next 24 hours. Figure 2 (forecast using this method) and Figure 3 (The corresponding actual distribution diagram) are the comparison diagrams of the forecast and the actual situation for this individual case.

[0069] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A medium- and short-term rainstorm forecasting method based on deep learning and EC numerical forecasting, characterized in that: The following steps are involved: Obtain the actual rainfall observation data and EC numerical model forecast data over the years; The EC numerical model forecast sample data set is constructed based on the actual rainfall observation data of previous years and the EC numerical model forecast data; Obtaining forecast model multi-channel data based on the EC numerical model forecast sample data set, and preprocessing the forecast model multi-channel data; An initial short- to medium-term rainstorm forecast model is constructed based on the Unet structure, and a model training loss function is constructed. The initial short- to medium-term rainstorm forecast model is trained based on the model training loss function and the preprocessed forecast model multi-channel data to obtain the final short- to medium-term rainstorm forecast model; based on the final short- to medium-term rainstorm forecast model, a grid forecast field of future precipitation is obtained.

2. The short- to medium-term rainstorm forecasting method based on deep learning and EC numerical forecasting according to claim 1, characterized in that: The actual rainfall observation data are the 24-hour cumulative precipitation data of the automatic weather station; the EC numerical model forecast data include the rainfall forecast fields of several ensemble members of the EC ensemble forecast and the U and V wind field data of multiple levels of EC fine grids.

3. The short- to medium-term rainstorm forecasting method based on deep learning and EC numerical forecasting according to claim 2, characterized in that: The process of constructing the EC numerical model forecast sample data set includes: Each starting time within the time range corresponding to the actual rainfall observation data is used as the starting time of precipitation in the next 24 hours, and the EC numerical model forecast data before the starting time is extracted; the EC numerical model forecast data is screened according to the precipitation forecast requirements, and after screening, it is combined with the precipitation to obtain the EC numerical model forecast sample data set, wherein the precipitation forecast requirements include the forecast range and weather type.

4. The short- to medium-term rainstorm forecasting method based on deep learning and EC numerical forecasting according to claim 1, characterized in that: The forecast model multi-channel data includes rainfall forecast field input channel data and wind field forecast field input channel data.

5. The short- to medium-term rainstorm forecasting method based on deep learning and EC numerical forecasting according to claim 4, characterized in that: The process of obtaining multi-channel data for the forecast model includes: The rainfall forecast fields of several ensemble members of the EC ensemble forecast are reconstructed and K member rainfall forecast fields are selected. The reconstruction process includes: sorting the forecast values ​​of several ensemble members corresponding to each grid point from large to small, selecting the first K members respectively, and obtaining the K member rainfall forecast fields corresponding to each grid point as the input channel data of the rainfall forecast field; based on the U and V wind field data at three levels of 500hPa, 850hPa and 925hPa, the wind speed and wind direction of the 12h and 36h forecast fields at each level are calculated respectively to obtain the input channel data of the wind field forecast field.

6. The short- to medium-term rainstorm forecasting method based on deep learning and EC numerical forecasting according to claim 1, characterized in that: The latitude and longitude range of the model input is determined based on the demand for forecast information, and the multi-channel data of the forecast model is scaled according to the latitude and longitude range.

7. The method for medium- and short-term rainstorm forecasting based on deep learning and EC numerical forecasting according to claim 1, characterized in that: Based on the Unet structure, the attention mechanism and the dilated convolution module are introduced to obtain the Unet variant as the initial medium- and short-term heavy rain forecast model.

8. The method for medium- and short-term rainstorm forecasting based on deep learning and EC numerical forecasting according to claim 1, characterized in that: The model training loss function is: Where loss represents the total loss; N is the total number of samples; w s is the weight of the sth sample, usually used to adjust the loss of the minority class; y s The true label of the sth sample; The predicted label for the sth sample.

Images