阅读说明：本技术 基于三维雷达回波的降水预测方法、装置、电子设备、存储介质 (Rainfall prediction method and device based on three-dimensional radar echo, electronic equipment and storage medium ) 是由 张国平 高金兵 王阔音 惠建忠 王曙东 匡秋明 于 2020-12-24 设计创作，主要内容包括：本申请提供一种基于三维雷达回波的降水预测方法、装置、电子设备、计算机可读存储介质,该方法包括：将实时采集的三维雷达回波输入已训练的雨强预测网络,获得所述雨强预测网络输出的预测降雨量；基于与连续区域的三维雷达回波对应的预测降雨量,构建预测雨强矩阵；根据前一时次的历史三维雷达回波和所述三维雷达回波,计算所述三维雷达回波的运动矢量；依据所述运动矢量对所述预测雨强矩阵进行外推处理,获得降雨预报数据。本申请方案可以充分利用了三维雷达回波中的信息,从而得到精确指示降雨量的降雨预报数据。(The application provides a rainfall prediction method, a rainfall prediction device, electronic equipment and a computer-readable storage medium based on three-dimensional radar echoes, wherein the method comprises the following steps: inputting the three-dimensional radar echo acquired in real time into a trained rainfall intensity prediction network to obtain the predicted rainfall output by the rainfall intensity prediction network; constructing a forecast rainfall matrix based on the forecast rainfall corresponding to the three-dimensional radar echo of the continuous area; calculating a motion vector of the three-dimensional radar echo according to the historical three-dimensional radar echo of the previous time and the three-dimensional radar echo; and carrying out extrapolation processing on the forecast rainfall matrix according to the motion vector to obtain rainfall forecast data. According to the scheme, the information in the three-dimensional radar echo can be fully utilized, so that rainfall forecast data for accurately indicating rainfall is obtained.)

1. A precipitation prediction method based on three-dimensional radar echo is characterized by comprising the following steps:

inputting the three-dimensional radar echo acquired in real time into a trained rainfall intensity prediction network to obtain the predicted rainfall output by the rainfall intensity prediction network;

constructing a forecast rainfall matrix based on the forecast rainfall corresponding to the three-dimensional radar echo of the continuous area;

calculating a motion vector of the three-dimensional radar echo according to the historical three-dimensional radar echo of the previous time and the three-dimensional radar echo;

and carrying out extrapolation processing on the forecast rainfall matrix according to the motion vector to obtain rainfall forecast data.

2. The method of claim 1, wherein prior to inputting the three-dimensional radar echo into the rain intensity prediction network, the method further comprises:

searching a preset network information table according to the acquired time information of the three-dimensional radar echo, and acquiring a network information table item matched with the time information;

and taking the rain intensity prediction network corresponding to the searched network information table item as the rain intensity prediction network of the three-dimensional radar echo to be input.

3. The method of claim 1 or 2, wherein the rain intensity prediction network comprises a plurality of rain intensity prediction subnetworks;

the method for inputting the three-dimensional radar echo acquired in real time into the trained rainfall prediction network to obtain the predicted rainfall output by the rainfall prediction network comprises the following steps:

respectively inputting the three-dimensional radar echo into each rain intensity prediction sub-network to obtain the sub-prediction precipitation output by the rain intensity prediction sub-network;

and carrying out fusion processing on the plurality of sub-prediction precipitation quantities according to a fusion model to obtain the prediction precipitation quantity.

4. The method of claim 3, wherein the rain intensity prediction sub-network is trained by:

inputting sample data in the sample data set into a machine learning algorithm to obtain a sample prediction rainfall output by the machine learning algorithm; the sample data is historical three-dimensional radar echo, and the actual rainfall is labeled with the sample data;

adjusting network parameters of the machine learning algorithm based on a difference between the sample predicted rainfall and the actual rainfall;

and repeating the processes until the machine learning algorithm is converged to obtain the rain intensity prediction subnetwork.

5. The method of claim 4, wherein after training a plurality of rain intensity prediction subnetworks in the rain intensity prediction network, the method further comprises:

respectively inputting the sample data into each rain intensity prediction sub-network to obtain the sample predicted rainfall output by each rain intensity prediction sub-network;

and fitting to obtain the fusion model according to the sample predicted rainfall corresponding to each rain intensity prediction sub-network and the actual rainfall corresponding to the sample data.

6. The method of claim 4, wherein prior to inputting the sample data into the machine learning algorithm, the method further comprises:

acquiring a large amount of historical three-dimensional radar echoes and actual rainfall corresponding to the historical three-dimensional radar echoes;

dividing the historical three-dimensional radar echoes according to the acquisition time of the historical three-dimensional radar echoes and preset time period information to obtain historical three-dimensional radar echoes corresponding to a plurality of time period information;

taking historical three-dimensional radar echoes corresponding to a plurality of time period information as sample data, and putting the sample data into a sample data set corresponding to the time period information; wherein the label of the sample data is the actual rainfall.

7. The method of claim 6, wherein prior to partitioning the historical three-dimensional radar returns, the method further comprises:

dividing historical three-dimensional radar echoes corresponding to actual rainfall according to a plurality of preset rainfall levels to obtain historical three-dimensional radar echoes corresponding to each rainfall level;

judging whether the number of historical three-dimensional radar echoes of any rainfall level exists, wherein the ratio of the total number of the historical three-dimensional radar echoes to the total number of the historical three-dimensional radar echoes is smaller than a preset ratio threshold;

if so, resampling the historical three-dimensional radar echoes of the rainfall level, and increasing the number of the historical three-dimensional radar echoes corresponding to the rainfall level.

8. A precipitation prediction device based on three-dimensional radar returns, comprising:

the input module is used for inputting the three-dimensional radar echo acquired in real time into the trained rainfall prediction network to obtain the predicted rainfall output by the rainfall prediction network;

the construction module is used for constructing a rainfall prediction matrix based on the rainfall prediction corresponding to the three-dimensional radar echo of the continuous area;

the calculation module is used for calculating a motion vector of the three-dimensional radar echo according to the historical three-dimensional radar echo at the previous time and the three-dimensional radar echo;

and the extrapolation module is used for extrapolating the forecast rainfall matrix according to the motion vector to obtain rainfall forecast data.

9. An electronic device, characterized in that the electronic device comprises:

a processor;

a memory for storing processor-executable instructions;

wherein the processor is configured to perform the method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that the storage medium stores a computer program executable by a processor to perform the method of any one of claims 1-7 for three-dimensional radar echo based precipitation prediction.

Technical Field

The application relates to the technical field of atmospheric science, in particular to a rainfall prediction method and device based on three-dimensional radar echoes, electronic equipment and a computer-readable storage medium.

Background

The Doppler weather radar can detect precipitation particles in cloud and has wide application in short-term precipitation forecast. A rainfall forecasting method based on a Doppler weather radar mainly comprises the following two steps: radar echo quantitative precipitation estimation and extrapolation prediction. The common radar echo quantitative rainfall estimation is mainly realized according to a Z-R relation method, and the radar echo basic reflectivity is calculated through a formula indicating the Z-R relation, so that the rainfall intensity is obtained.

A common scanning mode for weather radar in business is VCP21(Volume Cover Pattern) mode, and 9 elevation primary reflectivity data can be obtained for each Volume scan. The rainfall intensity can be calculated by inputting a formula indicating the Z-R relationship, typically using the lowest elevation base reflectance value or a blended scanning reflectance value blended from multiple layers of elevation. In the method, the three-dimensional radar echo needs to be converted into the two-dimensional radar echo, and information in the three-dimensional radar echo cannot be fully utilized. In addition, in the VCP21 mode, the radar body sweep can be completed in about 6 minutes, and at the same time, the ground rain gauge can realize the observation with the accuracy of 0.1 mm. The existing radar echo quantitative precipitation estimation calculates the rainfall intensity of one hour in unit time, so that the accuracy and the instantaneity of short-term precipitation forecast are poor.

Disclosure of Invention

An object of the embodiments of the present application is to provide a precipitation prediction method, device, electronic device, and computer-readable storage medium based on three-dimensional radar echo, for implementing precipitation prediction according to the three-dimensional radar echo.

In one aspect, the application provides a precipitation prediction method based on three-dimensional radar echo, including:

inputting the three-dimensional radar echo acquired in real time into a trained rainfall intensity prediction network to obtain the predicted rainfall output by the rainfall intensity prediction network;

constructing a forecast rainfall matrix based on the forecast rainfall corresponding to the three-dimensional radar echo of the continuous area;

calculating a motion vector of the three-dimensional radar echo according to the historical three-dimensional radar echo of the previous time and the three-dimensional radar echo;

and carrying out extrapolation processing on the forecast rainfall matrix according to the motion vector to obtain rainfall forecast data.

In an embodiment, before inputting the three-dimensional radar echo into the rain intensity prediction network, the method further comprises:

searching a preset network information table according to the acquired time information of the three-dimensional radar echo, and acquiring a network information table item matched with the time information;

and taking the rain intensity prediction network corresponding to the searched network information table item as the rain intensity prediction network of the three-dimensional radar echo to be input.

In an embodiment, the rain intensity prediction network comprises a plurality of rain intensity prediction subnetworks;

the method for inputting the three-dimensional radar echo acquired in real time into the trained rainfall prediction network to obtain the predicted rainfall output by the rainfall prediction network comprises the following steps:

respectively inputting the three-dimensional radar echo into each rain intensity prediction sub-network to obtain the sub-prediction precipitation output by the rain intensity prediction sub-network;

and carrying out fusion processing on the plurality of sub-prediction precipitation quantities according to a fusion model to obtain the prediction precipitation quantity.

In one embodiment, the rain intensity prediction subnetwork is trained by:

inputting sample data in the sample data set into a machine learning algorithm to obtain a sample prediction rainfall output by the machine learning algorithm; the sample data is historical three-dimensional radar echo, and the actual rainfall is labeled with the sample data;

adjusting network parameters of the machine learning algorithm based on a difference between the sample predicted rainfall and the actual rainfall;

and repeating the processes until the machine learning algorithm is converged to obtain the rain intensity prediction subnetwork.

In an embodiment, after training a plurality of rain intensity prediction subnetworks in the rain intensity prediction network, the method further includes:

respectively inputting the sample data into each rain intensity prediction sub-network to obtain the sample predicted rainfall output by each rain intensity prediction sub-network;

and fitting to obtain the fusion model according to the sample predicted rainfall corresponding to each rain intensity prediction sub-network and the actual rainfall corresponding to the sample data.

In an embodiment, prior to inputting the sample data into the machine learning model, the method further comprises:

acquiring a large amount of historical three-dimensional radar echoes and actual rainfall corresponding to the historical three-dimensional radar echoes;

dividing the historical three-dimensional radar echoes according to the acquisition time of the historical three-dimensional radar echoes and preset time period information to obtain historical three-dimensional radar echoes corresponding to a plurality of time period information;

taking historical three-dimensional radar echoes corresponding to a plurality of time period information as sample data, and putting the sample data into a sample data set corresponding to the time period information; wherein the label of the sample data is the actual rainfall.

In an embodiment, before dividing the historical three-dimensional radar returns, the method further comprises:

dividing historical three-dimensional radar echoes corresponding to actual rainfall according to a plurality of preset rainfall levels to obtain historical three-dimensional radar echoes corresponding to each rainfall level;

judging whether the number of historical three-dimensional radar echoes of any rainfall level exists, wherein the ratio of the total number of the historical three-dimensional radar echoes to the total number of the historical three-dimensional radar echoes is smaller than a preset ratio threshold;

if so, resampling the historical three-dimensional radar echoes of the rainfall level, and increasing the number of the historical three-dimensional radar echoes corresponding to the rainfall level.

On the other hand, this application still provides a rainfall prediction device based on three-dimensional radar echo, includes:

the input module is used for inputting the three-dimensional radar echo acquired in real time into the trained rainfall prediction network to obtain the predicted rainfall output by the rainfall prediction network;

the construction module is used for constructing a rainfall prediction matrix based on the rainfall prediction corresponding to the three-dimensional radar echo of the continuous area;

the calculation module is used for calculating a motion vector of the three-dimensional radar echo according to the historical three-dimensional radar echo at the previous time and the three-dimensional radar echo;

and the extrapolation module is used for extrapolating the forecast rainfall matrix according to the motion vector to obtain rainfall forecast data.

Further, the present application also provides an electronic device, including:

a processor;

a memory for storing processor-executable instructions;

wherein the processor is configured to perform the above three-dimensional radar echo based precipitation prediction method.

In addition, the present application also provides a computer-readable storage medium, which stores a computer program, which is executable by a processor to perform the above-mentioned three-dimensional radar echo based precipitation prediction method.

In the scheme, the rainfall prediction network is used for processing the three-dimensional radar echo acquired in real time, the predicted rainfall can be obtained, and a rainfall prediction matrix can be constructed according to the corresponding rainfall prediction of the continuous region; after a motion vector of the three-dimensional radar echo is calculated according to the three-dimensional radar echo and the historical three-dimensional radar echo at the previous time, extrapolation processing can be carried out on the forecast rainfall matrix according to the motion vector, and therefore rainfall forecast data are obtained;

the rainfall intensity prediction network can directly process the three-dimensional radar echo, and information in the three-dimensional radar echo is fully utilized, so that a rainfall intensity prediction matrix which accurately reflects rainfall is obtained; and after the rainfall prediction matrix is subjected to extrapolation processing through the running vector of the three-dimensional radar echo, rainfall forecast data which accurately indicate rainfall can be obtained.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required to be used in the embodiments of the present application will be briefly described below.

Fig. 1 is a schematic view of an application scenario of a three-dimensional radar echo-based precipitation prediction method according to an embodiment of the present application;

fig. 2 is a schematic structural diagram of an electronic device according to an embodiment of the present application;

fig. 3 is a schematic flowchart of a precipitation prediction method based on three-dimensional radar returns according to an embodiment of the present disclosure;

FIG. 4 is a schematic illustration of a region provided in accordance with an embodiment of the present application;

FIG. 5 is a flow chart illustrating training of a rain intensity prediction subnetwork according to an embodiment of the present application;

fig. 6 is a schematic flowchart of a sample data generation method according to an embodiment of the present application;

fig. 7 is a block diagram of a three-dimensional radar echo based precipitation prediction device according to an embodiment of the present application.

Detailed Description

The technical solutions in the embodiments of the present application will be described below with reference to the drawings in the embodiments of the present application.

Like reference numbers and letters refer to like items in the following figures, and thus, once an item is defined in one figure, it need not be further defined and explained in subsequent figures. Meanwhile, in the description of the present application, the terms "first", "second", and the like are used only for distinguishing the description, and are not to be construed as indicating or implying relative importance.

Fig. 1 is a schematic view of an application scenario of a precipitation prediction method based on three-dimensional radar returns according to an embodiment of the present application. As shown in fig. 1, the application scenario includes a client 20 and a server 30; the client 20 may be a computer, a mobile phone, a tablet or other devices of the radar weather station, and is configured to transmit radar echo data obtained by each physical scanning of a weather radar in the radar weather station to the server 30; the server 30 may obtain radar echo data from the clients 20 of the plurality of radar weather stations, thereby achieving precipitation prediction from the radar echo data.

As shown in fig. 2, the present embodiment provides an electronic apparatus 1 including: at least one processor 11 and a memory 12, one processor 11 being exemplified in fig. 2. The processor 11 and the memory 12 are connected by a bus 10, and the memory 12 stores instructions executable by the processor 11, and the instructions are executed by the processor 11, so that the electronic device 1 can execute all or part of the flow of the method in the embodiments described below. In an embodiment, the electronic device 1 may be the server 30, and is configured to perform a precipitation prediction method based on three-dimensional radar returns.

The Memory 12 may be implemented by any type of volatile or non-volatile Memory device or combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic Memory, flash Memory, magnetic disk or optical disk.

The present application also provides a computer readable storage medium storing a computer program executable by a processor 11 to perform the method for three-dimensional radar echo based precipitation prediction provided herein.

Referring to fig. 3, a flowchart of a method for predicting precipitation based on three-dimensional radar returns according to an embodiment of the present disclosure is shown in fig. 3, where the method may include the following steps 310 to 340.

Step 310: and inputting the three-dimensional radar echo acquired in real time into the trained rainfall prediction network to obtain the predicted rainfall output by the rainfall prediction network.

The rainfall intensity prediction network is used for calculating corresponding rainfall based on the three-dimensional radar echo data. The raininess prediction network can be obtained by training machine learning algorithms such as SVR (Support Vector Regression), ridge algorithm and the like.

The three-dimensional radar echo is reflectivity data obtained by volume scanning of a meteorological radar. The weather radar can obtain the reflectivity data of a plurality of elevation angles according to a preset scanning mode, and the reflectivity data of each elevation angle can be regarded as a matrix corresponding to a height plane, so that the reflectivity data of the elevation angles are superposed together, and a three-dimensional radar echo containing the matrix of a plurality of channels can be obtained. For example, the weather radar scans in a VCP21 mode to obtain reflectivity data of 9 elevation angles, and a three-dimensional radar echo containing 9 channel matrixes can be constructed based on the reflectivity data of the 9 elevation angles.

The meteorological radar can obtain the three-dimensional radar echo only after a certain time length is needed for each scanning. Illustratively, in VCP mode, each scan takes about 6 minutes to complete. The three-dimensional radar echo is reflectivity data of the meteorological radar in the time period from the scanning start to the scanning end. Thus, the three-dimensional radar echo corresponds to time information of the scanning period, which may be represented in a start time point and an end time point. For example, the weather radar starts scanning from 10 points and finishes scanning within 10 points and 6 minutes, and the time information can be 10 points to 10 points and 6 minutes.

Furthermore, the area covered by the weather radar can be determined by taking the position of the weather radar as the center. The three-dimensional radar returns correspond to position information for the area, which may be expressed in latitude and longitude coordinates. For example, the weather radar covers a rectangular area, and the position information corresponding to the three-dimensional radar returns may be longitude and latitude coordinates of four vertices of the rectangular area.

And after the three-dimensional radar echo corresponding to any time information and any position information is input into the rain intensity prediction network by the server, the predicted precipitation corresponding to the time information and the position information can be obtained.

Step 320: and constructing a forecast rainfall matrix based on the forecast rainfall corresponding to the three-dimensional radar echo of the continuous region.

The continuous area refers to a plurality of geographically continuous areas. Referring to fig. 4, a schematic region diagram is provided for an embodiment of the present application, as shown in fig. 4, a horizontal dotted line represents a weft, a vertical dotted line represents a warp, 8 regions divided by the warp and the weft in fig. 4 are respectively numbered 1001, 1002, 1003, 1004, 1005, 1006, 1007, and 1008, the region 1001 is geographically continuous with the region 1005 and the region 1002, and the region 1002 is geographically continuous with the region 1001, the region 1003, and the region 1006.

After the server side obtains the predicted rainfall corresponding to the multiple three-dimensional radar echoes in the continuous area, a predicted rainfall matrix can be constructed, and each element in the predicted rainfall matrix is the predicted rainfall. Taking fig. 4 as an example, the predicted rainfall capacity of 8 regions can be obtained, and a matrix with 2 rows and 4 columns is constructed, wherein the element in the 1 st row and the 1 st column in the matrix is the predicted rainfall capacity of the region 1001 in the continuous region, the element in the 1 st row and the 2 nd column in the matrix is the predicted rainfall capacity of the region 1002 in the continuous region, the element in the 1 st row and the 3 rd column in the matrix is the predicted rainfall capacity of the region 1003 in the continuous region, and so on.

Step 330: and calculating the motion vector of the three-dimensional radar echo according to the historical three-dimensional radar echo and the three-dimensional radar echo at the previous time.

The historical three-dimensional radar echo refers to a three-dimensional radar echo before the current three-dimensional radar echo is obtained.

The server can acquire and store the three-dimensional radar echo at each time. After the current three-dimensional radar echo is obtained, the server side can calculate the matrix of each channel in the previous historical three-dimensional radar echo and the matrix of the same channel in the three-dimensional radar echo according to an optical flow method, and therefore the motion vector of the three-dimensional radar echo is obtained. The motion vector may represent the motion of the precipitation particle. For example, the three-dimensional radar echo includes a matrix of 9 channels, and the server may calculate the matrix of each channel and the matrix of the same channel in the historical three-dimensional radar echo according to an optical flow method, so as to obtain vectors of 9 channels, and use the vectors of the 9 channels as motion vectors.

Step 340: and carrying out extrapolation processing on the forecast rainfall matrix according to the motion vector to obtain rainfall forecast data.

The server can select the vectors of a plurality of specified channels from the motion vectors to perform fusion processing. Here, the specified channel may be set empirically. For example, the server may select, from the motion vectors, vectors of multiple channels corresponding to a horizontal plane between 2000 meters and 4000 meters high above the ground for the fusion process. The server can fuse the vectors of the channels through averaging processing. During the averaging process, the server may calculate an average of elements at the same position in each channel vector, so as to obtain a fused motion vector.

The server side can perform extrapolation processing on the forecast rainfall matrix through the fused motion vector, so as to obtain rainfall forecast data. The rainfall forecast data indicates the rainfall capacity of the corresponding area of the forecast rainfall matrix in a specified time period in the future. Here, the specified time period may be determined by the acquisition time interval of two temporal three-dimensional radar echoes. For example, after the server calculates the predicted rainfall matrix according to the three-dimensional radar echoes collected between 10 points and 6 points of 8 areas in fig. 4, a motion vector is calculated according to the historical three-dimensional radar echoes collected between 9 points and 54 points and 10 points and the three-dimensional radar echoes, and the predicted rainfall matrix can be extrapolated according to the motion vector, so as to obtain rainfall forecast data which can represent the rainfall of the 8 areas in 10 points and 6 points and 10 points and 12 points.

For the same region, the rainfall indicated by the same three-dimensional radar echo also has difference in different months and seasons. In an embodiment, the server configures a raininess prediction network corresponding to a plurality of time periods, and configures a network information table. The network information table comprises a plurality of network information table entries, and each network information table entry can record the mapping relation between the time period information and the network identification of the rainfall intensity prediction network; the network identifier may be a number of the rain-strength prediction network or a storage location on the server for indicating the rain-strength prediction network.

Here, the period information may be a month, a season, a solar term, or the like. For example, a rain intensity prediction network corresponding to each month may be configured on the server, and the time period information is a name of the month.

Before the three-dimensional radar echo is input into the rain intensity prediction network, the server side can search the network information table according to the time information of the collected three-dimensional radar echo, and obtain a network information table item matched with the time information. Illustratively, the time period information in the network information table is month, and after the server obtains the three-dimensional radar echo at 10 o 'clock and 6 o' clock of 9 month, 7 day, 10 o 'clock, after the three-dimensional radar echo is obtained, the network information table can be searched according to the time information "10 o' clock and 6 o 'clock of 9 month, 7 day, 9 o' clock" to obtain the network information table item matched with the time information, where the time period information in the network information table item is "9 month".

The server can use the found rain intensity prediction network corresponding to the network information table item as the rain intensity prediction network of the three-dimensional radar echo to be input. The server may input the three-dimensional radar echo into the selected rain intensity prediction network when performing step 310.

In an embodiment, the rain intensity prediction network may include a plurality of rain intensity prediction subnetworks. In this case, when the server performs step 310, the three-dimensional radar echo may be respectively input into each rain intensity prediction sub-network, so as to obtain the sub-prediction precipitation output by each rain intensity prediction sub-network. Wherein, the rain intensity prediction sub-network can be obtained by training according to the same or different machine learning algorithms.

The server side can perform fusion processing on the plurality of sub-prediction precipitation quantities according to the fusion model to obtain the prediction precipitation quantity. For example, the fusion model may be a weighting formula including a weight corresponding to each rain intensity prediction sub-network. The server side can perform weighted summation on the plurality of sub-prediction rainfall amounts according to the fusion model, so that the prediction rainfall amount is obtained.

In an embodiment, before the foregoing precipitation prediction method is executed, the server may train a rain intensity prediction sub-network, see fig. 5, which is a training flowchart of the rain intensity prediction sub-network provided in an embodiment of the present application, and as shown in fig. 5, the training process may include the following steps 307 to 309.

Step 307: inputting sample data in the sample data set into a machine learning algorithm to obtain a sample prediction rainfall output by the machine learning algorithm; and the sample data is historical three-dimensional radar echo, and the actual rainfall is labeled with the sample data.

The sample data set includes a large number of historical three-dimensional radar returns as sample data. And the acquisition time of the historical three-dimensional radar echo in the sample data set is positioned in the same time period. For example, the raininess prediction network (raininess prediction sub-network) is divided by seasons, and then four seasons have corresponding sample data sets respectively.

The initial network parameters of the machine learning algorithm may be randomly generated. After the server side inputs the sample data into the machine learning algorithm, the predicted rainfall capacity of the sample corresponding to the sample data output by the machine learning algorithm can be obtained.

Step 308: and adjusting the network parameters of the machine learning algorithm based on the difference between the sample predicted rainfall and the actual rainfall.

Step 309: and repeating the processes until the machine learning algorithm is converged to obtain the rain intensity prediction subnetwork.

The server can evaluate the difference between the predicted rainfall of the sample and the actual rainfall marked on the sample data according to a preset loss function, and adjust the network parameters of the machine learning algorithm according to the difference. Through repeated iteration, when the function value of the loss function tends to be stable, the machine learning algorithm can be determined to be converged, and the rain intensity prediction sub-network is obtained. And if the rain intensity prediction network only comprises a unique rain intensity prediction sub-network, obtaining the rain intensity prediction network when the training of the rain intensity prediction sub-network is finished.

In an embodiment, after the server obtains the multiple rain intensity prediction subnetworks in the rain intensity prediction network through training, the server may input sample data into each rain intensity prediction subnetwork respectively to obtain a sample predicted rainfall output by each rain intensity prediction subnetwork. For example, after the server side trains a plurality of rain intensity prediction sub-networks according to a sample data set in summer, sample data in the sample data set can be respectively input into each rain intensity prediction sub-network, so that the predicted rainfall of the sample output by the rain intensity prediction sub-network is obtained.

The server side can obtain a fusion model through fitting according to the sample predicted rainfall amount corresponding to each rain intensity prediction sub-network and the actual rainfall amount corresponding to the sample data. The server can process the actual rainfall and the predicted rainfall of the samples corresponding to each actual rainfall according to a least square method, and therefore a fusion model is fitted.

In an embodiment, referring to fig. 6, a flowchart of a sample data generating method provided in an embodiment of the present application is shown in fig. 6, where the method may include the following steps 304 to 306.

Step 304: a number of historical three-dimensional radar returns and actual rainfall corresponding to the historical three-dimensional radar returns are obtained.

Wherein the actual rainfall corresponding to the historical three-dimensional radar echo is an actual rainfall corresponding to the same time information and the same position information as the historical three-dimensional radar echo. Here, the time information corresponding to the actual rainfall may be adjusted on a minute level according to actual needs, and in general, the actual rainfall may be a rainfall within 6 minutes.

Step 305: and dividing the historical three-dimensional radar echoes according to the acquisition time of the historical three-dimensional radar echoes and preset time period information to obtain the historical three-dimensional radar echoes corresponding to a plurality of time period information.

The time period information can be month, season, solar terms and the like, and can be adjusted according to application requirements.

After the server divides the historical three-dimensional radar echo, the historical three-dimensional radar echo corresponding to each time period information can be obtained. For example, the server may classify the acquired historical three-dimensional radar returns into historical three-dimensional radar returns corresponding to each month.

Step 306: taking historical three-dimensional radar echoes corresponding to the time period information as sample data, and putting the sample data into a sample data set corresponding to the time period information; wherein, the label of the sample data is the actual rainfall.

And after the server side takes the historical three-dimensional radar echo as sample data, marking the actual rainfall corresponding to the historical three-dimensional radar echo as a sample label, and putting the sample data into a corresponding sample data set. For example, the server may put the historical three-dimensional radar echo of 12 months into a sample data set corresponding to 12 months after adding the mark. And the subsequent server can train the rain intensity prediction sub-network corresponding to the time period information according to the sample data set. Since the time information corresponding to the actual rainfall amount of the sample label can indicate a short time period (which can be as short as several minutes), the prediction accuracy of the rain intensity prediction sub-network obtained by subsequent training is very high.

In an embodiment, after obtaining the historical three-dimensional radar echo, the server may also perform data preprocessing on the historical three-dimensional radar echo before dividing the historical three-dimensional radar echo. The server can divide the historical three-dimensional radar echo corresponding to the actual rainfall according to a plurality of preset rainfall levels to obtain the historical three-dimensional radar echo corresponding to each rainfall level. Wherein each rainfall level has a corresponding rainfall interval. And the server divides the historical three-dimensional radar echo into various rainfall levels according to the actual rainfall corresponding to the historical three-dimensional radar echo.

The server can judge whether the number of the historical three-dimensional radar echoes of any rainfall level exists, and the ratio of the total number of the historical three-dimensional radar echoes to the preset ratio threshold is smaller than the preset ratio threshold. The ratio threshold may be an empirical value, for example, there are 5 rainfall levels in total, and the ratio threshold may be set to 17% to ensure that the number of historical three-dimensional radar returns is close for each rainfall level.

When the number of the historical three-dimensional radar echoes of any rainfall level exists, and the ratio of the number of the historical three-dimensional radar echoes to the total number of the historical three-dimensional radar armguards is smaller than a ratio threshold, resampling processing can be performed on the historical three-dimensional radar echoes of the rainfall level, and therefore the number of the historical three-dimensional radar echoes corresponding to the rainfall level is increased. By the aid of the measures, the quantity of the sample data corresponding to each rainfall level can be guaranteed to be balanced, and accordingly training effects of subsequent rain intensity prediction sub-networks are improved.

Referring to fig. 7, a block diagram of a three-dimensional radar echo based precipitation prediction apparatus according to an embodiment of the present invention is shown in fig. 7, where the apparatus may include: an input module 710, a construction module 720, a calculation module 730, an extrapolation module 740.

The input module 710 is configured to input a three-dimensional radar echo acquired in real time into a trained rainfall prediction network, so as to obtain a predicted rainfall output by the rainfall prediction network;

a building module 720, configured to build a predicted rainfall matrix based on the predicted rainfall corresponding to the three-dimensional radar echo in the continuous region;

the calculating module 730 is configured to calculate a motion vector of the three-dimensional radar echo according to the historical three-dimensional radar echo of the previous time and the three-dimensional radar echo;

and the extrapolation module 740 is configured to perform extrapolation processing on the forecast rainfall matrix according to the motion vector to obtain rainfall forecast data.

The implementation processes of the functions and actions of each module in the device are specifically described in detail in the implementation processes of corresponding steps in the three-dimensional radar echo-based precipitation prediction method, and are not described herein again.

In the embodiments provided in the present application, the disclosed apparatus and method can be implemented in other ways. The apparatus embodiments described above are merely illustrative, and for example, the flowchart and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods and computer program products according to various embodiments of the present application. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). In some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.

In addition, functional modules in the embodiments of the present application may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.

The functions, if implemented in the form of software functional modules and sold or used as a stand-alone product, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present application or portions thereof that substantially contribute to the prior art may be embodied in the form of a software product stored in a storage medium and including instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method according to the embodiments of the present application. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk or an optical disk, and other various media capable of storing program codes.