CN116840941A - Model precipitation prediction correction method based on deep learning and cloud satellite - Google Patents

Abstract

The invention relates to a model precipitation prediction correction method based on deep learning and wind cloud satellites, which takes cloud image data of wind cloud satellites as input, builds a U-shaped precipitation prediction correction network combining a cavity convolution and an attention mechanism, improves the accuracy and reliability of 24-hour precipitation prediction, and comprises five layers of encoders and five layers of decoders corresponding to the five layers of encoders, wherein the encoders are used for extracting characteristic information related to future precipitation in the cloud image data, and the decoders receive the characteristic information related to precipitation, gradually reconstruct precipitation prediction information and finally output precipitation prediction correction results. The satellite remote sensing system has the advantages of fully playing the advantages of satellite remote sensing, avoiding the influence and limitation on the terrain complex area caused by the lack of station observation and weather radar observation data, along with simple structure and better timeliness.

Description

Model precipitation prediction correction method based on deep learning and cloud satellite

Technical Field

The invention relates to the field of model forecast correction, in particular to a model precipitation forecast correction method based on deep learning and cloud satellite.

Background

The numerical weather model forecast is a method for predicting the atmospheric movement trend and weather phenomenon in a certain time in the future by carrying out numerical calculation on the atmospheric state under certain boundary value and initial value conditions through a super computer and solving a thermodynamic and hydrodynamic equation set describing the weather change process, and is one of main means and tools for the current weather forecast. Because of the highly complex nonlinearity of atmospheric motion, the numerical weather forecast mode cannot accurately simulate the actual initial conditions, boundary conditions and complex evolution process of the atmosphere, certain errors inevitably exist in the mode forecast result, and the forecast error can be gradually increased along with the increase of forecast aging, so that the error is particularly obvious for the extreme weather process. In addition, the precipitation occurrence has the characteristics of high randomness, discontinuity and the like, and the difficulty of mode forecasting is further increased. The numerical weather forecast modes mainly adopted at home and abroad at present are as follows: a european mesoscale weather forecast model (ECMWF) developed by the middle european weather forecast center, a Global Forecast System (GFS) developed by the national environmental forecast center, a global and regional weather forecast model system (GRAPES) developed by the chinese weather agency, and the like. Among them, the ECMWF mode is one of the most advanced weather forecast modes worldwide.

The weather forecast with high precision has important guiding function and application value for a plurality of industries and fields such as disaster prevention and reduction, agricultural production, traffic and the like. However, precipitation occurrence has the characteristics of high nonlinearity, randomness, discontinuity and the like, and brings great challenges and difficulties to numerical weather pattern prediction, and the pattern precipitation prediction inevitably has errors. The model precipitation prediction order is to correct the precipitation prediction result of the digital weather prediction model by using the observed precipitation data so as to improve the accuracy and reliability of precipitation prediction. The conventional model correction method mainly comprises a model-based correction method and a statistical-based correction method. The model-based correction method mainly starts from mode prediction, improves the quality of initial field data, adjusts parameters of a numerical mode and the like to improve the accuracy of a prediction result. For example, the prediction accuracy of a pattern may be improved by optimizing the physical parameters of the pattern; the accuracy of the forecast may be improved by, for example, weighted averaging of the results of the multiple mode forecasts. The statistical-based correction method is relatively simple, mainly comprises the steps of statistically analyzing the difference between historical observation data and a numerical mode forecast output result, and then establishing models such as regression models, probability matching and the like for correction.

In recent years, inspired by successful application of artificial intelligence deep learning methods in numerous fields such as computer vision, natural language processing and the like, more and more students begin focusing on a prediction correction method based on deep learning. At present, a precipitation prediction correction research work based on deep learning is proposed, and the research result shows that the effectiveness, feasibility and advantages of the deep learning method in precipitation prediction correction are superior to those of the traditional method. However, existing precipitation prediction correction methods based on deep learning mostly rely on ground station observations and/or weather radar observations. However, due to the limitation of geographic conditions and maintenance cost, the problem of missing ground station observation data and weather radar observation data exists in a part of the terrain complex area, so that the expansion application of most of the existing rainfall forecast correction methods is limited. Compared with ground station observation data and weather radar observation data, satellite remote sensing has the advantages of continuous all-weather observation, wide coverage range and high time precision, and can provide weather data of continuous all-weather observation. However, at present, the research work of model precipitation forecast correction based on satellite remote sensing, especially the cloud satellite data of China is relatively less. The research of the model precipitation prediction correction method based on the remote sensing of the cloud satellite has a certain promotion effect on improving the accuracy and reliability of model precipitation prediction and improving the application value and influence of the cloud satellite in China in the global weather prediction business.

Precipitation is one of the most important meteorological elements, and has a vital effect on aspects of production and living, water resource utilization management, agricultural production management planning and the like of people. Therefore, the model is required to correct the water forecast result so as to improve the accuracy and reliability of the rainfall forecast, and further meet the application requirements.

Disclosure of Invention

Aiming at the defects of the prior art, a model precipitation prediction correction method based on deep learning and cloud satellites is provided, which is characterized in that cloud image data of the cloud satellites is taken as input, a U-shaped precipitation prediction correction network combining a cavity convolution and an attention mechanism is built, the U-shaped precipitation prediction correction network comprises five layers of encoders and five layers of decoders corresponding to the five layers of encoders, the encoders are used for extracting characteristic information related to future precipitation in the cloud image data, the decoders receive the characteristic information related to the precipitation, gradually reconstruct precipitation prediction information, and finally output precipitation prediction correction results, and the method specifically comprises the following steps:

step 1: collecting remote sensing cloud image data of a wind cloud satellite No. four, atmospheric re-analysis product data set ERA5 and ECMWF model 24-hour rainfall forecast data, preprocessing possible missing values and abnormal values in the data, selecting a research area, performing space-time matching processing, and unifying the data to the same area, the same time and the same spatial resolution;

step 2: dividing the data processed in the step 1 into a training set and a testing set according to a certain proportion, wherein longitude and latitude information of a selected area, satellite cloud image data and 24-hour rainfall forecast data in an ECMWF mode are used as inputs of a U-shaped rainfall forecast correction network, and an atmospheric analysis data set ERA5 is used as a true value to evaluate the accuracy of the rainfall forecast correction result;

step 3: satellite cloud image data, ECMWF mode 24-hour rainfall forecast data and longitude and latitude auxiliary variables at 4 moments are selected from a training set, and the satellite cloud image data, the ECMWF mode rainfall forecast data and the longitude and latitude auxiliary variables are spliced along the channel direction and subjected to [0,1] standardization processing to form data containing 27 channels;

step 4: inputting the output of the step 3 into a precipitation prediction correction network, and outputting a precipitation prediction correction result, wherein the precipitation prediction correction result is specifically as follows:

step 41: inputting the data processed in the step 3 into the encoder, and extracting characteristic information related to future precipitation from satellite cloud image data, wherein the method comprises the following steps:

step 411: the 1 st coding layer performs feature extraction on input data through two convolution operations to obtain first feature information X1, and outputs the first feature information X1 to the next coding layer network and an equal pair layer network of a decoder;

step 412: after receiving input information, the 2 nd coding layer firstly analyzes and strengthens useful information in the input information through channel attention and space attention; then carrying out downsampling operation on the processed information, then carrying out cavity convolution, conventional convolution, batch normalization and nonlinear activation operation which are the same as step 411 to obtain second characteristic information X2, and transmitting the second characteristic information X2 to a next coding layer network and a corresponding decoding layer network of a decoder;

step 413: repeatedly executing step 412 on the 3 rd coding layer, the 4 th coding layer and the 5 th coding layer to obtain third characteristic information X3, fourth characteristic information X4 and fifth characteristic information X5 respectively, and transmitting the fifth characteristic information X5 to a decoder;

step 42: inputting the characteristic information extracted by the coding layer into a corresponding decoding layer through jump connection to perform characteristic decoding, and finally outputting a precipitation forecast correction result, wherein the method comprises the following steps:

step 421: the 5 th decoding layer receives the fifth characteristic information X5 transmitted by the 5 th encoding layer and the fourth characteristic information X4 transmitted by the 4 th encoding layer, and firstly, deconvolution operation is carried out on the fifth characteristic information X5; then, the space size of the deconvolution operation result is adjusted to be consistent with the space size of the fourth characteristic information X4, then the deconvolution operation result is spliced with the fourth characteristic information X4 along the channel dimension, the two convolution operations are carried out on the spliced fusion result, and finally, the obtained fifth decoding information D5 is transmitted to a 4 th decoding layer network;

step 422: the 4 th decoding layer receives the fifth decoding information D5 transmitted by the 5 th decoding layer and the third characteristic information X3 transmitted by the encoder peer layer, and firstly, deconvolution operation is carried out on the fifth decoding information D5 transmitted by the next decoding layer; splicing the third characteristic information X3 transmitted by the encoder peer layer after the size adjustment along the channel dimension; then performing the same cavity convolution, conventional convolution, batch normalization and nonlinear activation operations as step 421; finally, the obtained fourth decoding result D4 is transmitted to a 3 rd decoding layer network;

step 423: repeatedly executing step 422 until the second decoding information D2 is obtained after the processing of the 2 nd coding layer, and transmitting the second decoding information D2 to the 1 st decoding layer;

step 424: after receiving the second decoding information D2, the 1 st decoding layer carries out 1X 1 convolution operation on the second decoding information D2 and outputs a precipitation prediction correction result, wherein the number of channels is 1, and the width and the height of the channel are consistent with the size of the cloud image input data;

step 5: the root mean square error and the average absolute error are used as a loss function, the difference between the rainfall forecast correction result output by the depth network and the true value is calculated, namely the loss value of the network is calculated, and then gradient back propagation and network parameter updating optimization are carried out based on the loss value;

step 6: judging whether the network training is converged, if not, continuing to execute the steps 3 to 5; if yes, the depth network parameters obtained through training are saved, and step 7 is executed;

step 7: the data in the test set are spliced along the channel direction according to satellite data, ECMWF model precipitation prediction data and longitude and latitude auxiliary variables at 4 moments, and are input into the trained precipitation prediction correction network after [0,1] standardization processing is carried out, and the correction network outputs corresponding precipitation prediction correction results;

step 8: and (3) statistically analyzing the difference between the precipitation forecast correction result of the depth network obtained on the test set and the ERA5 result, and evaluating the performance of the depth network correction model.

Compared with the prior art, the invention has the beneficial effects that:

1. according to the invention, through the combination of cavity convolution, an attention mechanism and a U-shaped deep network, the space and time dependency relationship between satellite observation data at different moments is deeply analyzed, so that the ECMWF model precipitation prediction result is corrected, and the accuracy and reliability of ECMWF model 24-hour precipitation prediction are improved.

2. The invention fully utilizes the current and historical wind cloud satellite cloud image data of the No. four A star (FY 4-A) at 4 moments, fully exerts the advantages of satellite remote sensing, such as continuous observation, wide coverage, high time precision and the like, and avoids the influence and limitation on a terrain complex area caused by station observation and weather radar observation data missing measurement.

3. Compared with the traditional method and other network architectures, the method has better timeliness and can achieve the effects of instant input and instant output.

Drawings

FIG. 1 is a schematic diagram of a precipitation prediction correction network according to the present invention;

FIG. 2 is a graph showing the result of comparative analysis of spatial distribution of precipitation prediction correlation coefficients for each method;

FIG. 3 is a graph showing the results of comparative analysis of the spatial distribution of root mean square errors in precipitation forecast for each method.

Detailed Description

The objects, technical solutions and advantages of the present invention will become more apparent by the following detailed description of the present invention with reference to the accompanying drawings. It should be understood that the description is only illustrative and is not intended to limit the scope of the invention. In addition, in the following description, descriptions of well-known structures and techniques are omitted so as not to unnecessarily obscure the present invention.

The following detailed description refers to the accompanying drawings.

Aiming at the defects of the prior art, the invention provides a model precipitation prediction correcting method based on deep learning and cloud satellite, which aims to correct ECMWF model precipitation prediction by using cloud satellite remote sensing data and improve the accuracy and reliability of ECMWF model precipitation prediction. Firstly, satellite remote sensing cloud image data of current and historical total 4 times are selected, and the 9 th to 14 th water vapor related spectrum channels of the satellite cloud image are selected at each time; cloud image data of 4 moments, regional longitude and latitude auxiliary variables and ECMWF mode forecast precipitation data of 24 hours in future with the same regional scope and spatial resolution are input into an encoder network after being spliced along the channel direction; the encoder comprises five network layers, and each network layer gradually extracts characteristic information related to future precipitation in input information through hole convolution operation, conventional convolution operation, space and channel attention mechanism, and transmits the final characteristic information to a decoder network after obtaining the final characteristic information; the decoder network comprises five decoding layers, each decoding layer combines the information extracted in the encoding stage and the information input in the next layer, gradually reconstructs precipitation prediction information through deconvolution and convolution operation, and finally outputs a precipitation prediction correction result, and the structural schematic diagram of the precipitation prediction correction network is shown in figure 1.

The invention takes cloud image data of a wind cloud meteorological satellite as input, builds a U-shaped rainfall forecast correction network combining a cavity convolution and an attention mechanism, wherein the U-shaped rainfall forecast correction network comprises five layers of encoders and five layers of decoders corresponding to the five layers of encoders, the encoders are used for extracting characteristic information related to future rainfall in the cloud image data, the decoders receive the characteristic information related to the rainfall, gradually reconstruct rainfall forecast information, and finally output rainfall forecast correction results, and the method specifically comprises the following steps:

step 1: and collecting remote sensing cloud image data of a wind cloud satellite No. four, atmospheric re-analysis product data set ERA5 and ECMWF model 24-hour rainfall forecast data, preprocessing possible missing values and abnormal values in the data, selecting a research area, performing space-time matching processing, and unifying the data to the same area, the same time and the same spatial resolution.

Step 2: dividing the data processed in the step 1 into a training set and a testing set according to a certain proportion, specifically, wherein longitude and latitude information of a selected area, satellite cloud image data and 24-hour rainfall forecast data in an ECMWF mode are used as inputs of a U-shaped rainfall forecast correction network, and an atmospheric analysis data set ERA5 is used as a true value to evaluate the accuracy of the rainfall forecast correction result.

Step 3: satellite cloud image data, ECMWF mode 24-hour rainfall forecast data and longitude and latitude auxiliary variables at 4 moments are selected from the training set, wherein the satellite cloud image data comprises 9 th to 14 th spectrum channel data, the satellite cloud image data, the ECMWF mode rainfall forecast data and the longitude and latitude auxiliary variables are spliced along the channel direction and subjected to [0,1] standardization processing, and data containing 27 channels are formed.

Step 4: inputting the output of the step 3 into a precipitation prediction correction network, and outputting a precipitation prediction correction result, wherein the precipitation prediction correction result is specifically as follows:

step 41: inputting the data processed in the step 3 into an encoder, and extracting characteristic information related to future precipitation in cloud image data, wherein the characteristic information comprises the following steps:

step 411: the 1 st coding layer performs feature extraction on input data through two convolution operations to obtain first feature information X1, and outputs the first feature information X1 to the next coding layer network and an equal pair layer network of a decoder.

Step 412: after receiving input information, the 2 nd coding layer firstly analyzes and strengthens useful information in the input information through channel attention and space attention; then carrying out downsampling operation on the processed information to reduce the size of the space; then performing the same hole convolution, conventional convolution, batch normalization and nonlinear activation operations as step 411; and finally, transmitting the second characteristic information X2 obtained by operation to a next coding layer network and a corresponding decoding layer network of a decoder.

Step 413: the 3 rd, 4 th and 5 th encoding layers repeatedly perform step 412 to obtain the third, fourth and fifth characteristic information X3, X4 and X5, respectively, and pass the fifth characteristic information X5 to the decoder network.

Step 42: inputting the characteristic information extracted by the coding layer into a corresponding decoding layer through jump connection to perform characteristic decoding, and finally outputting a precipitation forecast correction result, wherein the method comprises the following steps:

step 421: the 5 th decoding layer receives the fifth characteristic information X5 transmitted by the 5 th encoding layer and the fourth characteristic information X4 transmitted by the 4 th encoding layer, and firstly, deconvolution operation is carried out on the fifth characteristic information X5; and then, the space size of the deconvolution operation result is adjusted to be consistent with the space size of the fourth characteristic information X4, then, the deconvolution operation is carried out on the deconvolution operation result and the fourth characteristic information X4 along the channel dimension, and finally, the obtained fifth decoding information D5 is transmitted to a 4 th decoding layer network.

Step 422: the 4 th decoding layer receives the fifth decoding information D5 transmitted by the 5 th decoding layer and the third characteristic information X3 transmitted by the encoder peer layer, and firstly, deconvolution operation is carried out on the fifth decoding information D5 transmitted by the next decoding layer; splicing the third characteristic information X3 transmitted by the encoder peer layer after the size adjustment along the channel dimension; then performing the same cavity convolution, conventional convolution, batch normalization and nonlinear activation operations as step 421; and finally, transmitting the obtained fourth decoding result D4 to a 3 rd decoding layer network.

Step 423: step 422 is repeated until the second decoding information D2 is obtained after the processing of the 2 nd encoding layer, and is transferred to the 1 st decoding layer.

Step 424: after receiving the second decoding information D2, the 1 st decoding layer carries out 1×1 convolution operation on the second decoding information D2, and outputs a precipitation prediction correction result, wherein the number of channels is 1, and the width and the height of the channel are consistent with the size of cloud image input data.

Step 5: and calculating the difference between the rainfall forecast correction result and the true value output by the depth network by taking the root mean square error and the average absolute error as a loss function, namely calculating the loss value of the network, and then carrying out gradient back propagation and network parameter updating optimization based on the loss value.

Step 6: judging whether the network training is converged, if not, continuing to execute the steps 3 to 5; if yes, the depth network parameters obtained through training are saved, and step 7 is executed.

Step 7: and (3) splicing the data in the test set along the channel direction according to satellite data (9 th to 14 th spectral channel data) at 4 moments, ECMWF model precipitation prediction data and longitude and latitude auxiliary variables, carrying out [0,1] standardization processing, and inputting the spliced data into the trained precipitation prediction correction model, and outputting a corresponding precipitation prediction correction result by the correction model.

Step 8: and (3) carrying out statistical analysis on the difference between the precipitation prediction correction result of the depth network obtained from the test set and the ERA5 result, and evaluating the performance of the depth network correction model.

In order to further illustrate the performance of the method, the method is compared with the existing commonly used U-Net deep learning correction method and the European mesoscale weather forecast mode ECMWF rainfall forecast developed by the European metaphase weather forecast center, and the atmospheric re-analysis product ERA5 of the European metaphase weather forecast center is used as a true value. The evaluation indexes adopted by the invention are Root Mean Square Error (RMSE) (root mean square error, RMSE) and pearson correlation coefficient CC (pearson correlation coefficient, CC), wherein the smaller the RMSE value is, the better the RMSE value is, which indicates that the better the prediction effect is; the larger and better CC indicates the better prediction effect, and the statistical results are shown in Table 1.

TABLE 1 rainfall forecast comparison with ERA5 as true value

FIG. 2 is a graph showing the spatial distribution of correlation coefficients of precipitation forecast according to various methods. The EC mode of fig. 2 (a), i.e., the ECMWF mode, fig. 2 (a) is the ECMWF mode prediction result, fig. 2 (b) is the U-Net correction result, and fig. 2 (c) is the method correction result of the present invention. As can be seen from fig. 2, the inventive method shows a whiter color in most areas than the U-Net method and the ECMWF pattern, i.e. the inventive method has a relatively higher correlation coefficient value.

FIG. 3 shows the result of comparative analysis of the root mean square error spatial distribution of precipitation forecast in each method. The EC mode in fig. 3 (a), i.e., the ECMWF mode, fig. 3 (a) shows the result of prediction of the ECMWF mode, fig. 3 (b) shows the result of U-Net correction, and fig. 3 (c) shows the result of correction by the method of the present invention. As can be seen from fig. 3, the inventive method showed a darker color in most areas than the U-Net method and the ECMWF mode, i.e. the inventive method had a relatively lower root mean square error. The comparative analysis results of fig. 2 and fig. 3 prove that the effectiveness of the correction of the rainfall forecast by the method is beneficial to improving the accuracy and the reliability of the rainfall forecast result of the ECMWF mode.

It should be noted that the above-described embodiments are exemplary, and that a person skilled in the art, in light of the present disclosure, may devise various solutions that fall within the scope of the present disclosure and fall within the scope of the present disclosure. It should be understood by those skilled in the art that the present description and drawings are illustrative and not limiting to the claims. The scope of the invention is defined by the claims and their equivalents.

Claims (1)

1. The model precipitation prediction correction method based on the deep learning and the cloud satellite is characterized in that a U-shaped precipitation prediction correction network combining a cavity convolution and an attention mechanism is built by taking cloud image data of the cloud weather satellite as input, the U-shaped precipitation prediction correction network comprises five layers of encoders and five layers of decoders corresponding to the five layers of encoders, the encoders are used for extracting characteristic information related to future precipitation in the cloud image data of the satellite, the decoders receive the characteristic information related to precipitation, gradually reconstruct the precipitation prediction information and finally output precipitation prediction correction results, and the method specifically comprises the following steps:

step 1: collecting remote sensing cloud image data of a wind cloud satellite No. four, atmospheric re-analysis product data set ERA5 and ECMWF model 24-hour rainfall forecast data, preprocessing possible missing values and abnormal values in the data, selecting a research area, performing space-time matching processing, and unifying the data to the same area, the same time resolution and the same space resolution;

step 2: dividing the data processed in the step 1 into a training set and a testing set according to a certain proportion, wherein longitude and latitude information of a selected area, satellite cloud image data and 24-hour rainfall forecast data in an ECMWF mode are used as inputs of a U-shaped rainfall forecast correction network, and an atmospheric analysis data set ERA5 is used as a true value to evaluate the accuracy of the rainfall forecast correction result;

step 3: satellite cloud image data, ECMWF mode 24-hour rainfall forecast data and longitude and latitude auxiliary variables at 4 moments are selected from a training set, and the satellite cloud image data, the ECMWF mode rainfall forecast data and the longitude and latitude auxiliary variables are spliced along the channel direction and subjected to [0,1] standardization processing to form data containing 27 channels;

step 4: inputting the data processed in the step 3 into a constructed rainfall prediction correction network, and outputting a rainfall prediction correction result after feature extraction of an encoder and feature recovery of a decoder, wherein the rainfall prediction correction result is specifically as follows:

step 41: inputting the data processed in the step 3 into the encoder, and extracting characteristic information related to future precipitation from satellite cloud image data, wherein the method comprises the following steps:

step 411: the 1 st coding layer performs feature extraction on input data through two convolution operations to obtain first feature information X1, and outputs the first feature information X1 to the next coding layer network and an equal pair layer network of a decoder, and the 2 nd decoding layer is collected;

step 412: after receiving the input information, the 2 nd coding layer firstly analyzes and strengthens useful information in the input information through channel attention and space attention, then carries out downsampling operation on the processed information, then carries out cavity convolution, conventional convolution, batch normalization and nonlinear activation operation which are the same as step 411 to obtain second characteristic information X2, and transmits the second characteristic information X2 to the next coding layer network and the equal pair layer network of the decoder;

step 413: repeatedly executing step 412 on the 3 rd coding layer, the 4 th coding layer and the 5 th coding layer to obtain third characteristic information X3, fourth characteristic information X4 and fifth characteristic information X5 respectively, and transmitting the fifth characteristic information X5 to a decoder;

step 42: inputting the characteristic information extracted by the coding layer into a corresponding decoding layer through jump connection to perform characteristic decoding, and finally outputting a precipitation forecast correction result, wherein the method comprises the following steps:

step 421: the 5 th decoding layer receives the fifth characteristic information X5 transmitted by the 5 th encoding layer and the fourth characteristic information X4 transmitted by the 4 th encoding layer, and firstly, deconvolution operation is carried out on the fifth characteristic information X5; then, the space size of the deconvolution operation result is adjusted to be consistent with the space size of the fourth characteristic information X4, then the deconvolution operation result is spliced with the fourth characteristic information X4 along the channel dimension, the two convolution operations are carried out on the spliced fusion result, and finally, the obtained fifth decoding information D5 is transmitted to a 4 th decoding layer network;

step 422: the 4 th decoding layer receives the fifth decoding information D5 transmitted by the 5 th decoding layer and the third characteristic information X3 transmitted by the encoder peer layer, and firstly, deconvolution operation is carried out on the fifth decoding information D5 transmitted by the next decoding layer; splicing the third characteristic information X3 transmitted by the encoder peer layer after the size adjustment along the channel dimension; then performing the same cavity convolution, conventional convolution, batch normalization and nonlinear activation operations as step 421; finally, the obtained fourth decoding result D4 is transmitted to a 3 rd decoding layer network;

step 423: repeatedly executing step 422 until the second decoding information D2 is obtained after the processing of the 2 nd coding layer, and transmitting the second decoding information D2 to the 1 st decoding layer;

step 424: after receiving the second decoding information D2, the 1 st decoding layer carries out 1X 1 convolution operation on the second decoding information D2 and outputs a precipitation prediction correction result, wherein the number of channels is 1, and the width and the height of the channel are consistent with the size of the cloud image input data;

step 5: the root mean square error and the average absolute error are used as a loss function, the difference between the rainfall forecast correction result output by the depth network and the true value is calculated, namely the loss value of the network is calculated, and then gradient back propagation and network parameter updating optimization are carried out based on the loss value;

step 6: judging whether the network training is converged, if not, continuing to execute the steps 3 to 5; if yes, the depth network parameters obtained through training are saved, and step 7 is executed;

step 7: the data in the test set are spliced along the channel direction according to satellite data, ECMWF model precipitation prediction data and longitude and latitude auxiliary variables at 4 moments, and are input into the trained precipitation prediction correction network after [0,1] standardization processing is carried out, and the correction network outputs corresponding precipitation prediction correction results;

step 8: and (3) statistically analyzing the difference between the precipitation forecast correction result of the depth network obtained on the test set and the ERA5 result, and evaluating the performance of the depth network correction model.