A short-term precipitation prediction method seamlessly fusing deep learning and numerical mode

Abstract

The application discloses a short-term precipitation prediction method of seamless integration of deep learning and numerical mode, comprising the following steps: constructing a short-term precipitation initial prediction model based on a Vison Transformer network, and constructing a short-term precipitation revised prediction model of seamless integration of deep learning and numerical mode based on an improved U-Net network; collecting various meteorological data in real time, inputting the data into a trained short-term precipitation prediction optimal model after data preprocessing, and obtaining 0-6h short-term precipitation prediction; and after the 2-6h short-term precipitation prediction of a real-time regional numerical mode is fused, the short-term precipitation is input into a trained short-term precipitation optimal revised prediction model, and 0-6h seamless integration of short-term precipitation prediction is obtained. The application solves the problems of discontinuity in time, inconsistency in spatial position and low prediction accuracy in the deep learning extrapolation and numerical mode short-term precipitation fusion prediction in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to meteorological nowcasting precipitation prediction technology, in particular to a nowcasting precipitation prediction method based on seamless integration of deep learning and numerical model. BACKGROUND

[0002] Nowcasting precipitation prediction in 0-6h plays an important role in emergency management such as coastal agricultural production, urban safety and other decision-making relying on weather. Among them, nowcasting precipitation prediction technology based on deep learning can quickly obtain high-quality 0-2h precipitation prediction, but the prediction quality decreases rapidly after 2h. While nowcasting precipitation prediction technology based on numerical model is still in the initial start-up adjustment stage, and the corresponding prediction model is not stable, which cannot quickly obtain high-quality nonlinear precipitation prediction in 0-2h, but the prediction quality gradually stabilizes from 3h.

[0003] Numerical model is based on physical calculation, and the extrapolation prediction of deep learning belongs to statistical framework, which often has "spatial and temporal discontinuity" in the 3h prediction, that is, the precipitation prediction is not continuous in time and inconsistent in space. How to integrate nowcasting precipitation prediction based on numerical model and deep learning, improve the continuity and consistency in time and space, and improve the prediction accuracy has become a key problem. At present, researchers have tried to integrate the two predictions by assigning fixed weights in time, but the effect is not good, and there is still a big difference in spatial position, and the prediction accuracy is not high. SUMMARY

[0004] The purpose of the present application is to provide a nowcasting precipitation prediction method based on seamless integration of deep learning and numerical model, to solve the problem of discontinuity in time and inconsistency in space in the fusion prediction of deep learning extrapolation and numerical model nowcasting precipitation, and low prediction accuracy.

[0005] Technical scheme: In order to achieve the above purpose, the nowcasting precipitation prediction method based on seamless integration of deep learning and numerical model comprises the following steps:

[0006] Step 1, collect various meteorological data and perform data preprocessing, construct a nowcasting precipitation dataset in the form of "6h before tmin meteorological data-6h after tmin nowcasting precipitation real-time sample", and t represents the prediction interval time;

[0007] Step 2, based on the Vison Transformer network, construct a nowcasting precipitation initial prediction model, and train it using the nowcasting precipitation dataset to obtain a nowcasting precipitation prediction optimal model;

[0008] Step 3, re-collect and pre-process various meteorological data, input the short-term precipitation forecast optimal model, and obtain the initial short-term precipitation forecast of 0-6h per t min in the future;

[0009] Step 4, collect the 2-6h per t min short-term precipitation forecast of the regional numerical model consistent with the initial short-term precipitation forecast, and fuse with the initial short-term precipitation forecast, obtain the short-term precipitation fusion forecast, and further construct the short-term precipitation fusion forecast dataset in the form of "0-6h per t min short-term precipitation fusion forecast-0-6h per t min short-term precipitation real-time" sample form;

[0010] Step 5, based on the improved U-Net network, construct the short-term precipitation correction forecast model of seamless fusion of deep learning and numerical model, and train it using the short-term precipitation fusion forecast dataset to obtain the short-term precipitation forecast optimal correction forecast model;

[0011] Step 6, real-time collection of various meteorological data, data pre-processing, input into the short-term precipitation forecast optimal model, obtain the initial short-term precipitation forecast, and input into the short-term precipitation optimal correction forecast model after fusion with the real-time short-term precipitation forecast of the regional numerical model, obtain the 0-6h seamless fusion short-term precipitation forecast.

[0012] Preferably, the various meteorological data in step 1 include multi-element real-time analysis products and radar combined reflectivity factor network mosaics, and the multi-element real-time analysis products include 2m temperature, sea level pressure, 2m relative humidity, 10m wind field and precipitation.

[0013] Preferably, the data pre-processing includes replacing the abnormal values in the multi-element real-time analysis products with the annual climatic standard values, supplementing the missing data, using the inverse distance weighted interpolation method to process all meteorological data into grid fields with uniform spatio-temporal resolution, and normalizing the data of the unified grid field; The multi-element real-time analysis products and radar combined reflectivity factor network mosaics after normalization are constructed into a short-term precipitation dataset in the form of "first 6h per t min meteorological data-second 6h per t min short-term precipitation real-time".

[0014] Preferably, the short-term precipitation initial forecast model structure in step 2 includes a position coding input layer, a feature extraction layer and a decoding output layer, wherein the position coding input layer blocks and position codes the input various meteorological data to obtain a position coding feature vector, the feature extraction layer uses a multi-head attention mechanism to capture the spatio-temporal dependence between the position coding feature vectors, the decoding output layer adjusts the dimension and size of the feature vector output by the feature extraction layer, and maps to the normalized short-term precipitation forecast value on the future 0-6h per t min short-term precipitation real-time.

[0015] Preferably, the constructing the short-term precipitation fusion forecast dataset in step 4 comprises the following sub-steps:

[0016] Step 4.1, collecting short-term precipitation forecasts of the regional numerical model, with a forecast time of 2-6h per t min;

[0017] Step 4.2, weighting the short-term precipitation forecasts of the regional numerical model and the initial short-term precipitation forecasts to form a time-continuous short-term precipitation fusion forecast;

[0018] Step 4.3, based on the short-term precipitation fusion forecast and the multiple weather data in step 3, constructing a short-term precipitation fusion forecast dataset in the form of "0-6h per t min short-term precipitation fusion forecast - 0-6h per t min short-term precipitation live at the same time".

[0019] Preferably, the weighting method in step 4.2 is: the 2h precipitation forecast of the short-term precipitation forecasts of the regional numerical model and the initial short-term precipitation forecasts is given an initial forecast weight; as the forecast time increases, the forecast weight of the short-term precipitation forecasts of the regional numerical model is increased by an arithmetic difference, and the forecast weight of the initial short-term precipitation forecasts is decreased by an arithmetic difference; until the 6h precipitation forecast, the forecast weight of the short-term precipitation forecasts of the regional numerical model is increased to 1.0, and the forecast weight of the initial short-term precipitation forecasts is decreased to 0, to obtain a 0-6h time-continuous short-term precipitation fusion forecast.

[0020] Preferably, the improved U-Net network in step 5 is constructed based on the loss function of the Dice similarity coefficient score and the precipitation ETS score, and the loss function is:

[0021] ,

[0022] In the formula, , , respectively represent the precipitation ETS score loss term, the Dice similarity 0-6h forecast loss term and the Dice similarity 2-3h forecast loss term, , and respectively represent the weight coefficients of the precipitation ETS score loss term, the Dice similarity 0-6h forecast loss term and the Dice similarity 2-3h forecast loss term.

[0023] Preferably, the calculation formula is:

[0024] ,

[0025] ,

[0026] In the formula, represents the number of times of time, represents the number of times of false alarm, represents the number of times of missed alarm, represents the number of times of correct negative, represents the expected number of hits of random prediction.

[0027] Preferably, the calculation formula of the is as follows:

[0028] ,

[0029] In the formula, , respectively represent the short-range precipitation fusion forecast and the short-range precipitation actual situation of 0-6h per t min, , respectively represent the number of grid points of short-range precipitation fusion forecast and short-range precipitation actual situation respectively appearing precipitation, represent the number of grid points of short-range precipitation fusion forecast and short-range precipitation actual situation appearing precipitation at the same time;

[0030] The calculation formula of the is as follows:

[0031] ,

[0032] In the formula, , represent the adjacent 2 t min short-range precipitation fusion forecasts within 2-3h; , respectively represent the number of grid points of adjacent 2 t min short-range precipitation fusion forecasts respectively appearing precipitation within 2-3h; represent the weight coefficient of the s th Dice similarity score within 2-3h; represent the number of short-range precipitation forecasts within 2-3h.

[0033] Preferably, in the training process of the revised prediction model, the precipitation ETS score of the short-range precipitation fusion forecast, the Dice similarity of 0-6h prediction and the Dice similarity score of 2-3h prediction are continuously improved, the model is converged, and the improvement of the spatial consistency of the short-range precipitation fusion forecast of the 3rd prediction is achieved.

[0034] Beneficial effects: the present application has the following advantages: the present application firstly fuses two kinds of short-term precipitation forecasts based on deep learning and numerical mode in the prediction time, thereby improving the time continuity of the third hour forecast based on deep learning and numerical mode; the fused short-term precipitation fusion forecast is input into the short-term precipitation revision forecast model based on the improved U-Net network, the combination loss function of the Dice similarity coefficient and the precipitation ETS score is used to constrain the consistency of the short-term precipitation fusion forecast in space, so as to improve the spatial similarity of the third hour forecast of the short-term precipitation fusion forecast and improve the prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 It is a method flowchart of the present application;

[0036] Figure 2 It is a network structure diagram of the short-term precipitation initial forecast model;

[0037] Figure 3 It is a grid data color map of various meteorological data;

[0038] Figure 4 It is a network structure diagram of the short-term precipitation revision forecast model. DETAILED DESCRIPTION

[0039] The technical solutions of the present application will be described in detail below in combination with the embodiments and the drawings.

[0040] As Figure 1 shown, the short-term precipitation prediction method of the present application includes the following contents:

[0041] Step 1, collect various meteorological data in a historical time period and pre-process, and construct a short-term precipitation data set.

[0042] Among them, the various meteorological data includes: multi-element real-time analysis product and radar combined reflectivity factor network mosaic, the multi-element real-time analysis product includes: 2m air temperature, sea level pressure, 2m relative humidity, 10m wind field and precipitation 5 meteorological elements. Among them, the radar combined reflectivity factor can be understood as the structure characteristic of strong convective storm which can be directly reflected. The multi-element real-time analysis product can be understood as the direct representation of the thermal and dynamic factors of strong convective weather. The sea level pressure can be understood as the weight of the air column from the sea level to the upper boundary of the atmosphere per unit area.

[0043] The multi-element real-time analysis product combines the radar combined reflectivity factor network mosaic Figure 1 , and data preprocessing, specifically including the following steps:

[0044] Step 1.1, replace the abnormal value in the multi-element real-time analysis product with the annual climate state standard value (the average value in the past 30 years), supplement the missing data, and process all meteorological data into grid fields with unified spatial and temporal resolution by using the inverse distance weighted interpolation method to establish the dataset sample. The present application takes the grid field with a time resolution of 10 min and a horizontal spatial resolution of 1 km x 1 km as an example, and all the preprocessed and predicted output data are grid field data.

[0045] The inverse distance weighted interpolation method calculation formula is as follows:

[0046] (1),

[0047] (2),

[0048] In the formula, indicates the inverse distance weighted coefficient, indicates the Euclidean distance from the kth discrete site to the grid point to be interpolated, indicates the Euclidean distance from the kth discrete site to the grid point to be interpolated; indicates the number of discrete points around the interpolation point, indicates the kth discrete site around, indicates the kth discrete site, indicates the kth discrete site, indicates the kth discrete site, indicates the kth discrete site, indicates the kth discrete site,

[0049] Step 1.2, normalize the data of the unified grid field, convert them to the [0, 1] interval, and reduce the data variation range to improve the subsequent model training effect.

[0050] The normalization calculation formula is as follows:

[0051] (3),

[0052] In the formula, , respectively indicate the maximum value and the minimum value of the data in the dataset, indicates the normalized dataset sample, indicates any one grid point value in the dataset that needs to be normalized.

[0053] The normalized multi-element real-time analysis product and the radar combined reflectivity factor network mosaic are combined according to the sample form of "the first 6h 10 min meteorological data - the last 6h 10 min short-term precipitation forecast" to construct the short-term precipitation dataset.

[0054] Step 2, based on the Vison Transformer network, an initial short-term precipitation prediction model is established, and the short-term precipitation data set obtained in step 1 is used to train the initial short-term precipitation prediction model, and the optimal short-term precipitation prediction model based on the Vison Transformer network is obtained.

[0055] The short-term precipitation initial prediction model structure includes a position encoding input layer, a feature extraction layer, and a decoding output layer, as shown in Figure 2 .

[0056] The position encoding input layer: the input multiple meteorological data is blocked and position encoded to obtain a position encoding feature vector, replacing the traditional convolution operation to improve the feature extraction calculation efficiency. First, the blocking operation is performed, and then the position encoding is performed to obtain the position encoding feature vector after adding the encoding.

[0057] Blocking operation: the input multiple meteorological data is processed by non-overlapping blocking, and the multiple meteorological data is divided into N non-overlapping data blocks with a size of (16x16, for example):

[0058] (4),

[0059] N= (5),

[0060] In the formula, N represents the number of data blocks, represents the size of the data block, represents the input multiple meteorological data; represents the short-term precipitation data set obtained in step 1. The input multiple meteorological data contains four dimensions of channel number, latitude grid point number, longitude grid point number, and time step, which is input in the form of an equidistant latitude-longitude grid point matrix (for example: 8x256x256x36, which is the channel numberxlongitude grid point numberxlatitude grid point numberxtimes step), and the filled color plot of the grid point data is shown in Figure 3 .a is the 2-meter temperature grid point data filled color plot, Figure 3 b is the sea level pressure grid point data filled color plot, Figure 3 c is the 2-meter relative humidity grid point data filled color plot, Figure 3 d is the 10-meter wind field grid point data filled color plot, Figure 3 e is the 10-meter precipitation grid point data filled color plot, Figure 3 f is the radar combined reflectivity factor network mosaic grid point data filled color plot. Figure 3 The number of channels is represented, and in the present embodiment, the multi-element live analysis product (2-meter temperature, sea level pressure, 2-meter relative humidity, 10-meter wind field, and precipitation, each accounting for 1 channel) is 5 channels, and the radar combined reflectivity factor network mosaic has 3 channels, for a total of 8 channels. The number of latitude grid points is represented, The number of longitude grid points is represented, and S represents the time step.

[0061] The divided data block is represented as , The data form is a four-dimensional grid point data matrix, containing four dimensions of channel number C, latitude grid point number P, longitude grid point number P, and time step S (the number of latitude and longitude grid points is the size of the block data P x P), that is: .

[0062] Position encoding:

[0063] (6),

[0064] In the formula, The position encoding feature vector after adding the encoding is represented by position encoding and the data block to be encoded is added to obtain , which represents the data block to be encoded, The position encoding is assigned to each data block, and a learnable position encoding vector is obtained by automatic learning through back propagation in the initial prediction model training process of the short-term precipitation, in order to capture the spatial position information of the data block in the original data. The position encoding is added to the data block to be encoded to obtain the position encoding feature vector fused with position information. , The data form is a grid point data matrix, and the dimensions are consistent with , which are both four-dimensional grid point data matrices, that is: , .

[0065] Feature extraction layer: first, spatial and temporal feature projection is performed, the multi-head attention mechanism is used to capture the spatial and temporal dependence between position encoding feature vectors, then the data feature vectors of C channels are spliced and fused, and a regularization layer and standardization processing are added, and the feature vector is output.

[0066] Spatial and temporal feature projection: the high-dimensional and complex spatial and temporal features are decomposed into multiple low-dimensional features to improve the calculation efficiency and model expression ability, and the calculation formula is as follows:

[0067] (7),

[0068] In the formula, After the flattening operation, the position encoding feature vector converted into two dimensions, the first dimension is usually 1, and the second dimension is the original The multiplication of all dimensions, that is, , for example: the position encoding feature vector after adding the code 8x16x16x36, after the flattening operation, the two-dimensional position encoding feature vector is 1x73728; Indicates a learnable projection matrix, which is gradually optimized in the back propagation training of the short-term precipitation initial prediction model, and the data dimension is two-dimensional. The size of the first dimension is consistent with the second dimension of , and the second dimension G can be set as needed, represented as ; Indicates a query matrix, Indicates a key matrix, Indicates a value matrix, which is set by model parameters, and and, Determined, it can be ensured that The data dimension and size are consistent, that is, .

[0069] Further, using the multi-head attention mechanism to extract Global and local dependency relationships:

[0070] (8),

[0071] In the formula, Indicates a feature vector calculated by the multi-head attention mechanism, and the dimension of a single The dimension of a single is consistent with For example, the dimension and size of a single 1x2304, the dimension and size of a single Indicates a nonlinear function, Indicates a query matrix, Indicates a key matrix The transpose matrix of​​​​​​​​​​​​​ This represents the projection dimension of the query and key in each header.

[0072] The feature vectors of C channels are concatenated and fused, reducing the number of channels to 1. The calculation formula is as follows:

[0073] (9),

[0074] In the formula, This represents the feature vector after concatenation and fusion. Represents the weight matrix. Indicates the first to the second Feature vectors of multi-head attention mechanism for each data channel This indicates a linear splicing and merging operation.

[0075] The concatenated and merged feature vectors are processed through a regularization layer. Standardization is performed to eliminate internal covariate bias and prevent overfitting, vanishing gradients, or gradient explosion that could degrade generalization performance. The calculation formula is as follows:

[0076] (10)

[0077] In the formula, This represents the feature vector output by the regularization layer, with dimensions and size equal to... Consistency, that is G is set through parameters in the spatiotemporal feature projection step. , and , , The data dimensions and size are consistent. The normalization processing calculations for regularized layers typically include layer normalization, dropout methods, and residual chaining.

[0078] Decoding Output Layer: A decoder is built in the output layer, adjusting the feature vector output by the regularization layer. The dimensions and size are then mapped onto the 0-6h 10-minute precipitation data analysis product through a fully connected layer using a nonlinear transformation (normalized data). Finally, the data is denormalized to obtain the initial short-term precipitation forecast.

[0079] First, use matrix transformation Reshape() to transform the feature vectors. Feature reshaping is performed to obtain feature vectors with spatiotemporal structure. :

[0080] (11),

[0081] The eigenvectors obtained after matrix transformation Reshape() operation The dimensions are adjusted to five: batch size × number of longitudinal grid points × number of latitudinal grid points × time step × number of channels, taking 1×8×8×36×1 as an example, to facilitate subsequent network calculations. .

[0082] Furthermore, the feature vector is upsampled. The number of grid points in the longitudinal and latitudinal directions is adjusted to match the input meteorological data. The upsampling process is achieved through multi-layer transposed convolution, calculated as follows:

[0083] (12),

[0084] In the formula, This represents the feature vector output by the upsampling process; This indicates a transpose convolution operation; the specific number of layers depends on... The number of latitudinal and longitudinal grid points, as well as the convolution kernel size, determine the implementation in this embodiment. For example, we set up 5 transposed convolutional layers and one 4×4×1 convolutional kernel to perform 3D convolution. Each convolutional layer has a stride of 2×2×1, 1 input and 1 output channel, and 1×1×0 padding. The convolutional layers are stacked in a sequential cascade structure, meaning the output of each layer directly serves as the input to the next. The number of channels and the time step remain constant at 1 and 36, respectively. The feature vector... The size of varies across the five transposed convolutional layers as follows:

[0085] Transposed convolutional layer 1: Feature vector The change is from 1×8×8×36×1 to 1×16×16×36×1.

[0086] Transposed Convolutional Layer 2: Feature Vector The change is 1×16×16×36×1 → 1×32×32×36×1.

[0087] Transposed convolutional layer 3: Feature vector The change is 1×32×32×36×1 → 1×64×64×36×1.

[0088] Transposed convolutional layer 4: Feature vector The change is 1×64×64×36×1 → 1×128×128×36×1.

[0089] Transposed convolutional layer 5: Feature vector The result changes to 1×128×128×36×1 → 1×256×256×36×1.

[0090] In the fully connected layer, the feature vector output by the up-sampling process The flattened and nonlinearly transformed feature vector is mapped onto the future 0-6h 10-min precipitation nowcast product (after normalization) to obtain the normalized short-term precipitation forecast value, and the calculation formula is as follows:

[0091] (13),

[0092] In the formula, represents the flattened feature vector after the operation, is a two-dimensional data matrix, the first dimension is 1, and the second dimension is J , J= I x H x W x T x CO , that is, batch size × latitude grid point number × longitude grid point number × time step × output channel number. For example, 1 × 256 × 256 × 36 × 1, then the size of

[0093] (14),

[0094] In the formula, y represents the output precipitation normalized value, represents the bias term, represents the weight matrix of the feedforward neural network, () represents a nonlinear activation function, , , y are all two-dimensional data matrices, , L represents the size of the hidden layer corresponding to the weight matrix, which can be set according to the required network size, such as 2359296 J ) × 512 (L) ; for example, with a time step of 36, , that is the size of is L × 36 (time step), the size of

[0095] is 1 × L, 1 × 36, and the size of y is 1 × 36. The y is then denormalized, and the short-term precipitation forecast is output

[0096] :

[0097] In the formula, , respectively represent the maximum value and the minimum value in the data set sample, represents the precipitation normalized value.

[0098] In step 2, R, R h、R RF and R 1~ R 13 representing a set of different data matrices.

[0099] The short-term precipitation forecast optimal model training process is as follows: Adam is selected as the optimizer, the initial learning rate is set to lr=1e-4, the batch size is set according to the GPU memory, the maximum number of epochs is preset to max_epochs=20, the parameters are further adjusted for retraining according to the loss function decline, and the optimal model is obtained.

[0100] Step 3, re-collect and pre-process a plurality of meteorological data of at least one time series, input into the short-term precipitation forecast optimal model, and obtain an initial short-term precipitation forecast of 0-6h every 10 minutes in the future of at least one time series; wherein the re-collected plurality of meteorological data cannot have repetition or intersection with the short-term precipitation data set in step 1.

[0101] Step 4, collect the short-term precipitation forecast of the regional numerical model, and fuse it with the initial short-term precipitation forecast in step 3 to construct a short-term precipitation fusion forecast data set, which specifically includes the following steps:

[0102] Step 4.1, collect short-term precipitation forecast data of the regional numerical model (which specific regional numerical model can be determined according to the forecast area), the time series and interval time of the data are consistent with the initial short-term precipitation forecast, and the forecast time is taken as 2-6h every 10 minutes.

[0103] Step 4.2, the short-term precipitation forecast of the regional numerical model is weighted with the initial short-term precipitation forecast to form a time-continuous short-term precipitation fusion forecast, specifically:

[0104] The short-term precipitation forecast of the regional numerical model and the 2h precipitation forecast of the initial short-term precipitation forecast are assigned with an initial forecast weight, both of which are 0.5; with the increase of the forecast time, the forecast weight of the regional numerical model short-term precipitation forecast is increased by an arithmetic progression, and the forecast weight of the initial short-term precipitation forecast is decreased by an arithmetic progression; until the 6h forecast, the forecast weight of the regional numerical model short-term precipitation forecast is increased to 1.0, and the forecast weight of the initial short-term precipitation forecast is reduced to 0, to obtain a 0-6h time-continuous short-term precipitation fusion forecast, thereby preliminarily improving the time continuity of the numerical model and the 3h forecast based on the short-term precipitation forecast optimal model.

[0105] Step 4.3, combine the precipitation real-time data in the step 3 multiple weather data with the nowcasting precipitation fusion forecast in step 4.2, perform the step 1 preprocessing operation to obtain the grid field with unified space-time resolution, normalize the data to the interval [0, 1], and construct the nowcasting precipitation fusion forecast dataset in the form of "0-6h 10min nowcasting precipitation fusion forecast -> 0-6h 10min nowcasting precipitation real-time data at the same time".

[0106] Step 5, based on the improved U-Net network, construct a deep learning and numerical model seamless fusion nowcasting precipitation correction forecast model, and use the nowcasting precipitation fusion forecast dataset to train the nowcasting precipitation correction forecast model to obtain the optimal nowcasting precipitation forecast correction forecast model based on deep learning and numerical model seamless fusion, which includes the following steps:

[0107] Step 5.1, based on the loss function of Dice similarity coefficient score and precipitation ETS score, construct an improved U-Net network, and based on the improved U-Net network, construct a deep learning and numerical model seamless fusion nowcasting precipitation correction forecast model. The nowcasting precipitation correction forecast model structure is divided into three parts: downsampling, skip connection and upsampling. As shown in Figure 4 , the loss function is constructed based on the Dice similarity coefficient score and the precipitation ETS score to improve the spatial consistency and accuracy of the nowcasting precipitation fusion forecast, which is as follows:

[0108] Through the Dice similarity 0-6h prediction loss term, the spatial similarity between the nowcasting precipitation fusion forecast and the nowcasting precipitation real-time data is improved; through the Dice similarity 2-3h prediction loss term, the spatial consistency between the nowcasting precipitation fusion forecast 2-3h 10min is improved; through the precipitation ETS score loss term, the overall prediction accuracy of the nowcasting precipitation forecast is improved. In the training process of the nowcasting precipitation correction forecast model, the precipitation ETS score of the nowcasting precipitation fusion forecast, the Dice similarity of 0-6h prediction and the Dice similarity of 2-3h prediction are continuously improved, the model is converged, and the purpose of improving the spatial consistency of the nowcasting precipitation fusion forecast in the 3rd prediction and improving the overall prediction accuracy is achieved.

[0109] The loss function is as follows:

[0110] (16),

[0111] In the formula, represents the loss function of the nowcasting precipitation correction forecast model, , , respectively represent the precipitation ETS score loss term, the Dice similarity 0-6h prediction loss term and the Dice similarity 2-3h prediction loss term, , and respectively represent the weight coefficient of the precipitation ETS score loss term, the weight coefficient of the Dice similarity 0-6h prediction loss term and the weight coefficient of the Dice similarity 2-3h prediction loss term.

[0112] wherein, The calculation formula is:

[0113] (17),

[0114] (18),

[0115] In the formula, represents the time number of prediction accuracy, represents the event number of false alarm, represents the event number of missed alarm, represents the number of correct negative, represents the expected hit number of random prediction, and takes

[0116] The value range is: and is an integer. As can be seen from formula (17), RE is less than the minimum value, and at the same time, for the occurrence of precipitation event, L ETS , the denominator will not be zero.

[0117] The calculation formula of L

[0118] (19),

[0119] In the formula, , respectively represent the 0-6h 10min short-term precipitation fusion prediction and short-term precipitation actual situation, , respectively represent the grid number of precipitation occurrence of short-term precipitation fusion prediction and short-term precipitation actual situation, represents the grid number of precipitation occurrence of short-term precipitation fusion prediction and short-term precipitation actual situation at the same time.

[0120] The calculation formula of L

[0121] (20),

[0122] In the formula, , represent the 2-10min short-term precipitation fusion prediction of adjacent 2 10min within 2-3h; , respectively represent the number of grid points where the respective 10-minute short-term precipitation fusion forecast occurs within 2-3 hours; represent the weight coefficient of the s th Dice similarity score within 2-3 hours; represent the number of short-term precipitation forecasts within 2-3 hours. The present application takes 10-minute short-term precipitation forecasts as an example, so = 60min / 10min = 6.

[0123] Step 5.2, short-term precipitation correction forecast model In the training process, Adam is selected as the optimizer, appropriate initial learning rate, batch size and preset maximum number of rounds are set, and early stopping mechanism is combined to gradually adjust the loss function weight. First, preset: , , , to obtain the optimal model obtained by training.

[0124] Early stopping mechanism: if the validation loss does not decrease for 10 consecutive rounds, terminate the training, further adjust the loss function weight coefficient, learning rate, batch size and maximum number of rounds, and start training again until the loss function significantly decreases within 10 rounds.

[0125] Step 6, real-time collection of various meteorological data, after pre-processing in step 1, obtain the grid point field of various meteorological data with unified spatial and temporal resolution, input the short-term precipitation forecast optimal model based on the Vison Transformer network, and obtain the 0-6h 10-minute short-term precipitation forecast. The 0-6h 10-minute short-term precipitation forecast is weighted with the real-time regional numerical model 2-6h 10-minute short-term precipitation forecast to obtain a 0-6h time-continuous short-term precipitation fusion forecast. Then input the short-term precipitation fusion forecast into the short-term precipitation forecast optimal correction model based on deep learning and seamless integration of numerical model, obtain the 0-6h 10-minute normalized forecast value, and perform inverse normalization on the forecast value to obtain the deep learning and seamless integration of numerical model short-term precipitation forecast with time continuity, spatial position consistency and high prediction accuracy.

Claims

1. A method for short-term precipitation forecasting that seamlessly integrates deep learning and numerical models, characterized in that, Includes the following steps: Step 1: Collect various meteorological data and perform data preprocessing to construct a short-term precipitation dataset in the form of "6 hours of meteorological data at t min intervals before the forecast - 6 hours of short-term precipitation data at t min intervals after the forecast", where t represents the forecast interval. Step 2: Construct an initial short-term precipitation forecast model and train it using the short-term precipitation dataset to obtain the optimal short-term precipitation forecast model; Step 3: Collect and preprocess various meteorological data again, input them into the optimal short-term precipitation forecast model, and obtain the initial short-term precipitation forecast for 0-6 hours per minute. Step 4: Collect 2-6h t-min short-term precipitation forecasts from regional numerical models whose time series and time intervals are consistent with the initial short-term precipitation forecast, and fuse them with the initial short-term precipitation forecast to obtain a short-term precipitation fusion forecast. Further construct a short-term precipitation fusion forecast dataset in the form of "0-6h t-min short-term precipitation fusion forecast - 0-6h t-min short-term precipitation real-time data at the same time"; Step 5: Based on the improved U-Net network, construct a short-term precipitation correction forecast model that seamlessly integrates deep learning and numerical models, and train it using the short-term precipitation fusion forecast dataset to obtain the optimal short-term precipitation forecast correction model. Step 6: Collect multiple meteorological data in real time, input the preprocessed data into the optimal short-term precipitation forecast model to obtain the initial short-term precipitation forecast, and then input the fused short-term precipitation forecast with the real-time regional numerical model into the optimal corrected short-term precipitation forecast model to obtain a seamless fused short-term precipitation forecast from 0 to 6 hours. Based on the loss functions of Dice similarity coefficient score and precipitation ETS score, the improved U-Net network described in step 5 is constructed, with the loss function being: ， In the formula, , , These represent the ETS score loss term for precipitation, the Dice similarity 0-6h forecast loss term, and the Dice similarity 2-3h forecast loss term, respectively. , and These represent the weighting coefficients of the ETS precipitation score loss term, the weighting coefficients of the Dice similarity 0-6h forecast loss term, and the weighting coefficients of the Dice similarity 2-3h forecast loss term, respectively.

2. The short-term precipitation forecasting method based on seamless fusion of deep learning and numerical models according to claim 1, characterized in that, The various meteorological data mentioned in step 1 include multi-element real-time analysis products and radar combined reflectivity factor network mosaic. The multi-element real-time analysis products include: 2-meter air temperature, sea level air pressure, 2-meter relative humidity, 10-meter wind field, and precipitation.

3. The short-term precipitation forecasting method based on seamless fusion of deep learning and numerical models according to claim 2, characterized in that, The data preprocessing includes: replacing outliers in the multi-element real-time analysis products with annual climatological standard values ​​and supplementing missing data; then using inverse distance weighted interpolation to process all meteorological data into a grid field with uniform spatiotemporal resolution; normalizing the data in the uniform grid field; and then networking and mosaicking the normalized multi-element real-time analysis products and radar combined reflectivity factors, constructing a short-term precipitation dataset according to the sample format of "6 hours of meteorological data per minute before the first 6 hours - 6 hours of short-term precipitation data per minute after the last 6 hours".

4. The short-term precipitation forecasting method based on seamless fusion of deep learning and numerical models according to claim 1, characterized in that, The short-term precipitation initial forecast model structure described in step 2 includes: a location encoding input layer, a feature extraction layer, and a decoding output layer. The location encoding input layer divides and positions various input meteorological data to obtain location encoding feature vectors. The feature extraction layer captures the spatiotemporal dependencies between location encoding feature vectors. The decoding output layer adjusts the dimension and size of the feature vectors output by the feature extraction layer and maps them to the actual short-term precipitation conditions in the next 0-6 hours, obtaining normalized short-term precipitation forecast values.

5. The short-term precipitation forecasting method based on seamless fusion of deep learning and numerical models according to claim 1, characterized in that, Constructing the short-term precipitation fusion forecast dataset described in step 4 includes the following sub-steps: Step 4.1: Collect short-term precipitation forecasts from regional numerical models, with forecast times ranging from 2 to 6 hours per minute. Step 4.2: Weight the regional numerical model short-term precipitation forecast with the initial short-term precipitation forecast to form a time-continuous short-term precipitation fusion forecast; Step 4.3: Based on the fusion forecast of multiple meteorological data and short-term precipitation in Step 3, construct a short-term precipitation fusion forecast dataset according to the sample format of "0-6h t-min short-term precipitation fusion forecast - 0-6h t-min short-term precipitation real-time data at the same time".

6. The short-term precipitation forecasting method based on seamless fusion of deep learning and numerical models according to claim 5, characterized in that, The weighting method described in step 4.2 is as follows: assign the initial forecast weight to the 2-hour precipitation forecast of the regional numerical model short-term precipitation forecast and the initial short-term precipitation forecast; As the forecast time increases, the forecast weight of the short-term precipitation forecast in the regional numerical model is increased by an arithmetic progression, while the forecast weight of the initial short-term precipitation forecast is decreased by an arithmetic progression. Until the 6-hour precipitation forecast, the forecast weight of the short-term precipitation forecast in the regional numerical model increases to 1.0, and the forecast weight of the initial short-term precipitation forecast decreases to 0, thus obtaining a continuous short-term precipitation fusion forecast over the time span of 0-6 hours.

7. The short-term precipitation forecasting method based on seamless fusion of deep learning and numerical models according to claim 1, characterized in that, The The calculation formula is: ， ， In the formula, The number of times that indicate the accuracy of the forecast. Indicates the number of events that are false alarms. Indicates the number of events that were missed. Indicates the number of correct negations. This represents the expected number of hits for a random forecast.

8. The short-term precipitation forecasting method based on seamless fusion of deep learning and numerical models according to claim 1, characterized in that, The The calculation formula is: ， In the formula, , These represent the fused forecast and actual short-term precipitation data for 0-6 hours, t / min, respectively. , These represent the number of grid points where precipitation occurred in the short-term precipitation fusion forecast and the actual short-term precipitation, respectively. This indicates the number of grid points where precipitation occurs simultaneously in both the short-term precipitation fusion forecast and the actual short-term precipitation. The The calculation formula is: ， In the formula, , This indicates the fusion forecast of two adjacent short-term precipitation events within 2-3 hours; , These represent the number of grid points where precipitation occurred in each of the two adjacent t-min short-term precipitation fusion forecasts within 2-3 hours; This represents the weighting coefficient of the s-th Dice similarity score within 2-3 hours; This indicates the number of short-term precipitation forecasts within 2-3 hours.

9. The short-term precipitation forecasting method based on seamless fusion of deep learning and numerical models according to claim 1, characterized in that, During the training process of the revised forecast model, the precipitation ETS score of the short-term precipitation fusion forecast, the Dice similarity score of the 0-6h forecast, and the Dice similarity score of the 2-3h forecast are continuously improved to enhance the spatial consistency of the 3h forecast of the short-term precipitation fusion forecast.

Images