A summer precipitation grading prediction method, device and storage device

Abstract

The application provides a summer precipitation grading prediction method, device and storage device, reanalysis data of statistically significant prediction factors of summer precipitation in China and precipitation observation data in China are obtained; the obtained data is preprocessed, input data is composed of a plurality of prediction factors in the early stage, and the category of summer precipitation in the eastern region of China is taken as a target value; samples are made according to the input data and the target value, and all samples are divided into a training set, a validation set and a test set; the AttentionU-Net model is pre-trained using the training set and the validation set; the early stopping technology is used, the pre-training process is repeated, the training set and the validation set loss function are balanced, and the optimal AttentionU-Net model is selected; the test set samples are predicted to obtain actual precipitation grading prediction results. The application has the beneficial effects that the prediction accuracy is high, especially the accuracy and the correlation coefficient are obviously improved, and overfitting is effectively avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of precipitation classification prediction, in particular to a summer precipitation classification prediction method, device and storage device. BACKGROUND

[0002] At present, the method of seasonal precipitation prediction mainly depends on numerical model. The dynamic numerical model is based on physical empirical equation, and models the weather process. The main source of model error is the physical parameterization scheme of the model, which is due to the limitation of human professional knowledge in understanding the physical process of precipitation. This leads to the low skill of numerical model in seasonal prediction. The common error correction means is to correct the predicted value of the numerical model by statistical method as post-processing method. However, due to the highly complex nonlinear characteristics of precipitation, only dynamic prediction and existing error correction cannot meet the demand of seasonal precipitation prediction. With the gradual maturity of artificial intelligence technology, many deep learning seasonal prediction models have been proposed in recent years. However, these models have two problems of noise interference and overfitting which need to be solved.

[0003] The existing regression model cannot avoid the influence of extreme precipitation. The number of extreme precipitation is small and the intensity is large, and its accuracy is small, which can mislead the optimization of the model. In addition, overfitting is the result caused by the complexity of the model relative to the characteristics of the data, and overfitting can be solved from two aspects of model and data. The measures of model include increasing dropout layer. Data can use higher quality and more quantity of numerical model data to pretrain the model, and then use transfer learning. However, the implementation of dropout layer is relatively complex. If too many neurons are discarded, the characteristics cannot be completely learned, and even the model cannot converge. However, if too many neurons are retained, the effect of reducing the complexity of the model cannot be achieved. Transfer learning can effectively avoid the problem of overfitting, but the prediction distribution of dynamic numerical model is different from the observation data. This will cause a certain deviation in the prediction result. SUMMARY

[0004] In view of the above problems, the present application provides a summer precipitation classification prediction method, device and storage device based on attention mechanism and U-Net, which not only uses deep learning network to predict summer precipitation classification, avoids the influence of extreme precipitation on the model, and also uses principal component analysis (PC China East summer precipitation classification prediction) to reduce model parameters and early stopping technology to completely solve the problem of model overfitting. While improving the prediction accuracy, some technical methods for predicting seasonal precipitation by deep learning are also improved. A summer precipitation classification prediction method is proposed for the climate of China east, which mainly includes:

[0005] S1: obtaining reanalysis data of prediction factors with statistical significance of summer precipitation in China and China regional precipitation observation data;

[0006] S2: preprocessing the obtained data, input data is composed of a plurality of prediction factors in the early stage, and the category of summer precipitation in the eastern region of China is taken as the target value;

[0007] S3: preparing samples according to the input data and the target value, and dividing all the samples into three parts of a training set, a validation set and a test set;

[0008] S4: establishing an Attention U-Net model;

[0009] S5: pre-training the Attention U-Net model using the training set and the validation set;

[0010] S6: using the early stopping technique, repeating the step S5 process, and selecting a better pre-training model;

[0011] S7: using the pre-trained Attention U-Net model to predict the test set data, and when the set accuracy is reached, obtaining the final Attention U-Net model for predicting the actual precipitation classification.

[0012] Further, the reanalysis data in step S1 includes: 0-1 meter deep soil moisture and 2 meter high air temperature in China; 500 / 850 / 1000 hPa potential height in the northern hemisphere; 250 / 500 / 850 / 1000 hPa potential height in the north and south polar circles; North Pacific, North Indian Ocean sea surface temperature and 5 meter deep sea surface salinity; sea surface temperature of the North Atlantic; snow depth, snow temperature and snow albedo of the Qinghai-Tibet Plateau snow cover.

[0013] Further, the specific process of the preprocessing in step S2 is:

[0014] S2.1: the observed precipitation value is processed into precipitation anomaly percentage according to formula (1):

[0015]

[0016] Wherein, m represents the climate state, pre represents the precipitation, the unit is millimeter / day, pp is the precipitation anomaly percentage, the unit is %;

[0017] S2.2: taking the category of summer precipitation in the eastern region of China as the target value: the precipitation anomaly percentage in the eastern region of China is divided into 8 categories according to the numerical value, and the range of the 8 categories of precipitation is as follows: the first category: missing value; the second category: -100-45; the third category: -45-30; the fourth category: -30-15; the fifth category: -15-0; the sixth category: 0-18; the seventh category: 18-45; the eighth category: >45; taking these precipitation categories as the model target value, the influence of the few and strong extreme value noise on the model optimization is eliminated;

[0018] S2.3: The input data is composed of multiple predictors in the previous period: the input data specifically includes observed precipitation, air temperature and soil moisture in the previous 1-3 months; the potential height of the South and North Poles in spring; the potential height of the Northern Hemisphere, sea surface temperature, sea surface salinity, snow depth, snow temperature and snow albedo in the previous winter and spring; these variables are cut into the same size as the observed precipitation data, and then each variable is normalized according to the Z-Score, and the function definition of Z-Score is as follows:

[0019]

[0020] Where x and respectively represent the original data and the mean value of the original data, and s is the standard deviation;

[0021] S2.4: Superimpose multiple input data according to the channel dimension.

[0022] Further, the establishment process of the Attention U-Net model in step S4 is as follows:

[0023] S4.1: Principal component analysis is performed on the input data A, and the first two principal components are finally obtained as the input data A1 of the Attention U-Net model;

[0024] S4.2: The input data A1 is encoded by the encoder to obtain A5;

[0025] S4.3: The model output B is obtained by decoding A5 using the decoder.

[0026] Further, the operation of the encoder includes:

[0027] S4.2.1: The input data A1 is processed by the DoubleConv layer to obtain A2, the DoubleConv layer includes: two 3*3 convolutions, BatchNorm for normalization and Relu activation function, and the definition of BatchNorm is as follows:

[0028]

[0029] Where y is the normalized data, x is the original data, E[x] and Var[x] are the mean and variance calculated in the current iteration, respectively, and γ, β are the weight and bias, which are updated after each backpropagation;

[0030] The definition of the Relu activation function is as follows:

[0031] f(x) = max(0, x) (4)

[0032] where f(x) is the output of the activation function, and x is the input of the activation function.

[0033] S4.2.2: A3 is obtained after down-sampling processing of A2, the down-sampling comprising: a max-pooling layer and a DoubleConv layer, the following down-sampling comprising the same operation;

[0034] S4.2.3: A4 is obtained after down-sampling processing of A3.

[0035] S4.2.4: A5 is obtained after down-sampling processing of A4.

[0036] Further, the operation of the decoder comprises:

[0037] S4.3.1: B1 is obtained after up-sampling one processing of A5, the up-sampling one being a deconvolution layer;

[0038] S4.3.2: C1 is obtained after AG processing of B1 and A4, the AG comprising: an Attention Gage, the operation of the Attention Gage being defined as:

[0039] I input data encoding layer X and decoding layer G;

[0040] II, X1 and G1 are obtained after 1*1 convolution and BatchNorm processing of X and G, respectively;

[0041] III, the sum of X1 and G1 is calculated through a Relu activation function;

[0042] IV, the attention coefficient is obtained through 1*1 convolution, BatchNorm and Sigmoid activation function The definition of the Sigmoid activation function is as follows:

[0043]

[0044] where f(x) is the output of the activation function, and x is the input of the activation function.

[0045] V, output:

[0046] S4.3.3: B11 is obtained by concatenating B1 and C1 in the channel dimension;

[0047] S4.3.4: B2 is obtained after B11 is processed by up-sampling two, the up-sampling two comprising: a DoubleConv layer and a deconvolution layer;

[0048] S4.3.5: C2 is obtained after AG processing of B2 and A3.

[0049] S4.3.6: B2 and C2 are spliced according to the channel dimension to obtain B21;

[0050] S4.3.7: B21 is processed by upsampling two to obtain B3;

[0051] S4.3.8: B3 and A2 are processed by AG to obtain C3;

[0052] S4.3.9: B3 and C3 are spliced according to the channel dimension to obtain B31;

[0053] S4.3.10: B31 is processed by DoubleConv to obtain B4;

[0054] S4.3.11: B4 is processed by a Conv layer to obtain model output B, and the Conv layer is a 1*1 convolution layer.

[0055] Further, in step S5, the cross-entropy loss function is selected as the loss function of the model classification task, and the hyperparameters of the model are adjusted through the convergence of the loss function. The definition of the cross-entropy loss function torch.nn.CrossEntropyLoss is as follows:

[0056] torch.nn.CrossEntropyLoss = Softmax + ln + NLL_Loss (6)

[0057] ln = log e (Soft max) (8)

[0058]

[0059] Where Softmax represents the activation function, ln represents the logarithmic function, and NLL_Loss represents the negative log-likelihood loss; y i,j,k,l represents four-dimensional input data, C is the number of categories, and j represents the jth category of precipitation.

[0060] Further, in step S6, the early stopping method is used to reduce the number of iterations of the model and retain the best generalization ability of the model.

[0061] A storage device stores instructions and data for implementing a summer precipitation classification prediction method.

[0062] A summer precipitation classification prediction device includes a processor and the storage device; the processor loads and executes the instructions and data in the storage device to implement a summer precipitation classification prediction method.

[0063] The technical scheme provided by the present application brings the beneficial effects that:

[0064] 1、 Based on the existing research, the prediction factors are determined, and the prediction factors are subjected to PCA to reduce the model complexity, and the Attention U-Net model is applied to the prediction of the precipitation type. The advantages of U-Net suitable for small sample tasks and the attention module capable of automatically capturing the key area are utilized. The model has high prediction accuracy, especially in the accuracy and correlation coefficient.

[0065] 2、 The number of grid points of each precipitation type is uniform, which not only avoids the influence of uneven samples on the model, but also further eliminates the influence of extreme values on the optimization of the model.

[0066] 3、 The present application reduces the model parameters (reduces the depth of U-Net, only retains three layers of upsampling and downsampling), adopts principal component analysis and early stopping method to optimize the model, and effectively avoids overfitting. BRIEF DESCRIPTION OF DRAWINGS

[0067] The present application will be further described below in combination with the drawings and examples, and the drawings are as follows:

[0068] Figure 1 is a flow chart of a summer precipitation classification prediction method in an embodiment of the present application.

[0069] Figure 2 is a structure diagram of an Attention U-Net model in an embodiment of the present application.

[0070] Figure 3 is an early stopping technology diagram in an embodiment of the present application (since there are individual maximum values of the loss function, in order to more clearly show the minimum value, only the maximum value is taken to 2).

[0071] Figure 4 is a deep learning model prediction effect diagram in an embodiment of the present application.

[0072] Figure 5 is a schematic diagram of the working of the hardware device in an embodiment of the present application. DETAILED DESCRIPTION

[0073] In order to have a clearer understanding of the technical features, objects and effects of the present application, the specific embodiments of the present application will be described in detail with reference to the drawings.

[0074] The embodiment of the present application provides a summer precipitation classification prediction method, device and storage device. The previous researches are all using regression models to predict seasonal precipitation, but the present application uses a classification model for prediction.

[0075] Please refer to Figure 1 ,Figure 1 is a flow chart of a summer precipitation classification prediction method in an embodiment of the present application, and specifically comprises:

[0076] S1: Data preparation: obtain reanalysis data of statistically significant prediction factors of summer precipitation in China (including: 0-1 meter deep soil moisture and 2 meter high air temperature in China; 500 / 850 / 1000 hPa geopotential height in the Northern Hemisphere; 250 / 500 / 850 / 1000 hPa geopotential height in the South and North Poles; North Pacific, North Indian Ocean sea surface temperature and 5 meter deep sea surface salinity; sea surface temperature of the North Atlantic; snow depth, snow temperature and snow albedo of the Qinghai-Tibet Plateau snow); and precipitation observation data in China.

[0077] S2: Data preprocessing:

[0078] S2.1: Process the precipitation observation value into precipitation anomaly percentage according to formula (1), and the calculation of the anomaly percentage is as follows:

[0079]

[0080] Wherein, pre is the precipitation, the unit is millimeter / day; m is the climate state, the climate state uses the average value of 1991-2020; pp is the precipitation anomaly percentage, the unit is %.

[0081] S2.2: Taking the summer precipitation categories in the eastern region of China as target values, the precipitation anomaly percentages in the eastern region of China are divided into 8 categories according to the numerical size. The 8 precipitation ranges are as follows: the first category: missing value; the second category: -100-45; the third category: -45-30; the fourth category: -30-15; the fifth category: -15-0; the sixth category: 0-18; the seventh category: 18-45; the eighth category: >45. The masks of the 8 categories are 0-7 respectively. Except for the missing value, the total number of grid points in each category is about 150,000, and the maximum difference is not more than 20,000. In this way, the number of grid points in each category can be kept flat, and the model will not be biased towards the training of a certain category. Taking these summer precipitation categories in the eastern region of China as model target values can eliminate the influence of extreme value noise with small quantity and large intensity on model optimization.

[0082] S2.3: The input data is composed of multiple prediction factors in the previous period. Including: observed precipitation, air temperature and soil moisture 1-3 months in advance; the geopotential height of the South and North Poles in spring; the geopotential height of the Northern Hemisphere, sea surface temperature, sea surface salinity, snow depth, snow temperature and snow albedo in the previous winter and spring. These variables are cut into the same size as the observed precipitation data, i.e. high*wide=163*283. Then normalize each variable according to Z-Score. The function definition of Z-Score is as follows:

[0083]

[0084] where x and represent the original data and the mean of the original data, and s is the standard deviation.

[0085] S2.4: The multiple variables are stacked according to the channel dimension, that is, the size of the input data is 45*163*283.

[0086] S3: Sample making: a sample includes input and target value. All samples are divided into three parts of training set, validation set and test set. Their sample quantity ratio is 6:2:2. The samples of each stage are independent.

[0087] S4: Attention U-Net model building: the model structure diagram can be seen Figure 2 .

[0088] S4.1: Principal component analysis (PCA) is performed on the input data, and finally the first two principal components are obtained as model input, and the cumulative contribution rate of the two principal components is more than 90%. Principal component analysis can effectively reduce the complexity of the model.

[0089] S4.2: An encoder is used to encode A1 to obtain A5, and the operation of the encoder includes:

[0090] S4.2.1: A1 is processed by DoubleConv to obtain A2, and the DoubleConv layer includes: twice 3*3 convolution, BatchNorm for normalization and Relu activation function. The following DoubleConv includes the same operation. The BatchNorm and Relu function are defined as follows:

[0091]

[0092] where y is the normalized data, x is the original data, E[x] and Var[x] are the mean and variance calculated in the current iteration, respectively, and γ, β are the weight and bias, which are updated after each back propagation.

[0093] The definition of the Relu activation function is as follows:

[0094] f(x) = max(0, x) (4)

[0095] where f(x) is the output of the activation function, and x is the input of the activation function.

[0096] S4.2.2: A2 is processed by downsampling to obtain A3, and the downsampling includes: a max pooling layer and a DoubleConv layer. The following downsampling includes the same operation;

[0097] S4.2.3: A3 is processed by downsampling to obtain A4;

[0098] S4.2.4: A4 is processed by downsampling to obtain A5;

[0099] S4.3: A5 is decoded by a decoder to obtain model output B;

[0100] The operation of the decoder includes:

[0101] S4.3.1: A5 is processed by upsampling to obtain B1, and the upsampling is a deconvolution layer;

[0102] S4.3.2: B1 and A4 are processed by AG to obtain C1, and the AG includes: Attention Gage, and the operation of Attention Gage is defined as:

[0103] I input data: X (encoding layer) and G (decoding layer);

[0104] II After 1*1 convolution and BatchNorm processing on X and G, X1 and G1 are obtained;

[0105] III The sum of X1 and G1 is calculated through a Relu activation function;

[0106] IV The attention coefficient is obtained through 1*1 convolution, BatchNorm and Sigmoid activation function The Sigmoid activation function is defined as follows:

[0107]

[0108] Wherein, f(x) is the output of the activation function, and x is the input of the activation function.

[0109] V output:

[0110] S4.3.3: B1 and C1 are spliced according to the channel dimension to obtain B11;

[0111] S4.3.4: B11 is processed by the second upsampling to obtain B2, and the second upsampling includes: DoubleConv layer and deconvolution layer;

[0112] S4.3.5: B2 and A3 are processed by AG to obtain C2, and the AG operation is the same as above;

[0113] S4.3.6: B2 and C2 are spliced according to the channel dimension to obtain B21;

[0114] S4.3.7: B21 is processed by the second upsampling to obtain B3;

[0115] S4.3.8: C3 is obtained by processing B3 and A2 through AG, where the AG operation is consistent with the above;

[0116] S4.3.9: B3 and C3 are spliced according to the channel dimension to obtain B31;

[0117] S4.3.10: B31 is processed through DoubleConv (including two convolution layers) to obtain B4;

[0118] S4.3.11: B4 is processed through a Conv layer to obtain a model output B, where the Conv layer is a 1*1 convolution layer.

[0119] S5: Pre-training of Attention U-Net using the training set and the validation set. Cross-entropy loss function is selected as the loss function of the classification task. And the hyperparameters are adjusted through the convergence of the loss function. The cross-entropy loss function is defined as follows:

[0120] torch.nn.CrossEntropyLoss = Soft max + ln + NLL_Loss (6)

[0121]

[0122] ln = log e (Soft max) (8)

[0123]

[0124] Where Softmax represents the activation function, ln represents the logarithmic function, and NLL Loss represents the negative log-likelihood loss. The cross-entropy function is composed of these three algorithms. y i，j，k，l represents the input data of four dimensions, and the input data yi, j, k, l is calculated in three steps to obtain NLL_Loss representing the loss function. In formula 6, the four dimensions of the variable yi, j, k, l are batch_size, channel, height and width. C is the number of categories, and j represents the jth category of precipitation. Formula 7 is the calculation of the logarithm with base e. Finally, NLL_Loss is calculated.

[0125] S6: Use the early stopping technique to repeat the process of step S5 and select a better pre-trained model. As shown in Figure 3 the validation set loss function of the model with epoch = 100 reaches the lowest point and shows signs of rising. But when the epoch is reduced to 50, the model converges faster, the validation set loss function is lower, and the model also has better generalization ability.

[0126] S7: using the pre-trained model to predict the test set data. As shown in Figure 4 The summer precipitation model established by the application has high prediction accuracy and is superior to the traditional numerical model CFSv2. In particular, the accuracy and correlation coefficient are significantly improved.

[0127] Please refer to Figure 5 , Figure 5 is a hardware device working schematic diagram of an embodiment of the application, which specifically comprises: a summer precipitation grading prediction device 401, a processor 402 and a storage device 403.

[0128] The summer precipitation grading prediction device 401: the summer precipitation grading prediction device 401 realizes the summer precipitation grading prediction method.

[0129] The processor 402: the processor 402 loads and executes the instructions and data in the storage device 403 to realize the summer precipitation grading prediction method.

[0130] The storage device 403: the storage device 403 stores instructions and data; the storage device 403 is used to realize the summer precipitation grading prediction method.

[0131] The beneficial effects of the application are:

[0132] 1. Based on the existing research, the prediction factors are determined, and the PCA is performed on the prediction factors to reduce the model complexity, and the Attention U-Net model is applied to the prediction of the precipitation type. The advantages of U-Net suitable for small sample tasks and attention module capable of automatically capturing key areas are utilized. The model has high prediction accuracy, and in particular, the accuracy and correlation coefficient are significantly improved.

[0133] 2. The number of grid points of each precipitation type is uniform, which not only avoids the influence of uneven samples on the model, but also further eliminates the influence of extreme values on the optimization of the model.

[0134] 3. The deep learning seasonal prediction model mentioned above has the defect of overfitting. The application reduces the model parameters (reduces the depth of U-Net, only retains three layers of upsampling and downsampling), adopts principal component analysis and early stopping method to optimize the model, and effectively avoids overfitting.

[0135] The above only describes the preferred embodiments of the application and does not limit the application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the application shall be included in the protection scope of the application.

Claims

1. A method of graded prediction of summer precipitation, characterized by: The method comprises the following steps: S1: obtaining reanalysis data of statistically significant prediction factors of summer precipitation in China and precipitation observation data in China; S2: preprocessing the obtained data, forming input data from multiple prediction factors in the early stage, and taking the category of summer precipitation in the eastern region of China as the target value; S3: preparing samples according to the input data and the target value, and dividing all the samples into a training set, a validation set and a test set; S4: establishing an Attention U-Net model; The specific process is as follows: S4.1: performing principal component analysis on the input data A, and finally obtaining the first two principal components as the input data A1 of the Attention U-Net model; S4.2: encoding the input data A1 by using an encoder to obtain A5; S4.3: decoding A5 by using a decoder to obtain model output B; The operation of the encoder comprises: S4.2.1: processing the input data A1 through a DoubleConv layer to obtain A2, wherein the DoubleConv layer comprises two 3*3 convolution operations, a BatchNorm for normalization and a Relu activation function, and the BatchNorm is defined as follows: wherein y is the normalized data, x is the original data, E[x] and Var[x] are the mean and variance obtained by iteration, and γ and β are the weight and bias, which are updated after each back propagation; The definition of the Relu activation function is as follows: f(x) = max(0, x) (4) wherein f(x) is the output of the activation function, and x is the input of the activation function; S4.2.2: obtaining A3 by performing down-sampling on A2, wherein the down-sampling comprises a max-pooling layer and a DoubleConv layer, and the following down-sampling comprises the same operation; S4.2.3: obtaining A4 by performing down-sampling on A3; S4.2.4: obtaining A5 by performing down-sampling on A4; The operation of the decoder comprises: S4.3.1: obtaining B1 by performing up-sampling one on A5, wherein the up-sampling one is a deconvolution layer; S4.3.2: obtaining C1 by performing AG on B1 and A4, wherein the AG comprises an Attention Gage, and the operation of the Attention Gage is defined as follows: I input data encoding layer X and decoding layer G; II performing 1*1 convolution and BatchNorm processing on X and G to obtain X1 and G1; III calculating the sum of X1 and G1 through a Relu activation function; IV The attention coefficients are obtained by 1 * 1 convolution, BatchNorm and Sigmoid activation function The definition of the Sigmoid activation function is as follows: wherein f(x) is the output of the activation function, and x is the input of the activation function; Vout: S4.3.3: splicing B1 and C1 according to the channel dimension to obtain B11; S4.3.4: obtaining B2 by performing up-sampling two on B11, wherein the up-sampling two comprises a DoubleConv layer and a deconvolution layer; S4.3.5: obtaining C2 by performing AG on B2 and A3; S4.3.6: splicing B2 and C2 according to the channel dimension to obtain B21; S4.3.7: obtaining B3 by performing up-sampling two on B21. S4.3.8: C3 is obtained by processing B3 and A2 through AG; S4.3.9: B31 is obtained by splicing B3 and C3 according to the channel dimension; S4.3.10: B4 is obtained by processing B31 through DoubleConv; S4.3.11: B is obtained by processing B4 through a Conv layer, which is a 1*1 convolution layer; S5: The Attention U-Net model is pre-trained using the training set and the validation set; S6: The early stopping technique is used to repeat the process of step S5, and a better pre-trained model is selected; S7: The test set samples and the optimal Attention U-Net model obtained in S6 are used to predict the actual precipitation classification.

2. A method of graded prediction of summer precipitation as claimed in claim 1 wherein: The reanalysis data in step S1 includes: 0-1 meter deep soil moisture and 2 meter high air temperature in China; 500 / 850 / 1000 hPa geopotential height in the Northern Hemisphere; 250 / 500 / 850 / 1000 hPa geopotential height in the Arctic and Antarctic circles; North Pacific, North Indian Ocean sea surface temperature and 5 meter deep sea surface salinity; North Atlantic sea surface temperature; Qinghai-Tibet Plateau snow depth, snow temperature and snow albedo.

3. A method of graded prediction of summer precipitation as claimed in claim 1 wherein: The specific process of the preprocessing in step S2 is as follows: S2.1: The observed precipitation is processed into precipitation anomaly percentage according to formula (1): Where m represents the climate state, pre represents the precipitation, the unit is millimeter / day, and pp is the precipitation anomaly percentage, the unit is %; S2.2: The summer precipitation classification in eastern China is used as the target value: the precipitation anomaly percentage in eastern China is divided into 8 categories according to the numerical value, and the 8 categories of precipitation range are as follows: the first category: missing value; The second category: -100-45; the third category: -45-30; the fourth category: -30-15; the fifth category: -15-0; the sixth category: 0-18; the seventh category: 18-45; the eighth category: >45; these precipitation classifications are used as model target values to eliminate the influence of extreme value noise with small quantity and large intensity on model optimization; S2.3: The input data is composed of multiple prediction factors in the previous period: the input data specifically includes: observed precipitation, air temperature and soil moisture 1-3 months in advance; spring geopotential height in the Arctic and Antarctic circles; Northern Hemisphere geopotential height, sea surface temperature, sea surface salinity, snow depth, snow temperature and snow albedo in the previous winter and spring; these variables are cut into the same size as the observed precipitation data, and then normalized according to Z-Score, and the function definition of Z-Score is as follows: where x and respectively represent the original data and the mean value of the original data, and s is the standard deviation; S2.4: Superimpose multiple input data according to the channel dimension.

4. The summer rainfall grading prediction method of claim 1, wherein: In step S5, the cross-entropy loss function is selected as the loss function for the model classification task, and the hyperparameters of the model are adjusted through the convergence of the loss function. The definition of cross-entropy loss function torch.nn.CrossEntropyLoss is as follows: torch.nn.CrossEntropyLoss = Soft max + ln + NLL_Loss (6) ln = log e (Soft max) (8) where Softmax represents an activation function, ln represents a logarithm function, and NLL_Loss represents a negative log-likelihood loss; y i,j,k,l represents input data of four dimensions, C is the number of categories, and j represents the jth category of precipitation.

5. A method of graded prediction of summer precipitation as claimed in claim 1 wherein: In step S6, the early stopping method is used to reduce the iteration number of the model, and the best generalization ability of the model is reserved.

6. A storage device, characterized by: The storage device stores instructions and data for implementing the summer precipitation classification prediction method in any one of claims 1-5.

7. A summer rainfall grading prediction device characterized by comprising: Comprise: A processor and a storage device; the processor loads and executes instructions and data in the storage device to implement the summer precipitation classification prediction method in any one of claims 1-5.

Images