MF-SPCM-based sub-seasonal rainfall forecast correction method

Abstract

The invention discloses a sub-seasonal rainfall forecast correction method based on MF-SPCM, and relates to the technical field of rainfall forecast correction, and the method comprises the following steps: obtaining an S2S sub-seasonal rainfall forecast result, ERA5 historical meteorological data, digital elevation model (DEM) data, and north polar region vortex related data of a target area in the future 1-6 weeks provided by an NCEP; preprocessing the data to obtain digital elevation model (DEM) features, spatio-temporal features in the early stage of forecasting and a mapping result consistent with forecasting data dimensions, and extracting the spatio-temporal features; obtaining a feature map of the multi-source data; inputting the multi-source data feature map and rainfall forecast data to be corrected into a multi-source cross attention fusion module to obtain a weighted multi-source data feature map; obtaining the correction result of the second-season accumulated rainfall forecast in the future 1-6 weeks; the precision of sub-season rainfall forecast is effectively improved, and the timeliness is prolonged.

Description

Technical Field

[0001] The present invention belongs to the technical field of precipitation forecast correction, and particularly relates to a sub-seasonal precipitation forecast correction method based on MF-SPCM. Background Art

[0002] Sub-seasonal meteorological forecasting lies between weather forecasting and seasonal prediction. Since it receives less attention, it is also known as the "prediction desert". Sub-seasonal forecasting also has a strong guiding and prompting effect on industries closely related to weather and climate conditions such as power dispatching, fishery production, and agricultural development. However, the current sub-seasonal precipitation forecast still has a large deviation and is not sufficient for actual forecasting operations. Therefore, there is an urgent need to develop a method to improve the accuracy of precipitation forecasting at the sub-seasonal scale. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a sub-seasonal precipitation forecast correction method based on MF-SPCM in view of the deficiencies of the background art, effectively improving the accuracy of sub-seasonal precipitation forecasting and extending the timeliness.

[0004] The present invention adopts the following technical solutions to solve the above technical problems:

[0005] A sub-seasonal precipitation forecast correction method based on MF-SPCM specifically includes the following steps:

[0006] Step 1: Obtain the sub-seasonal precipitation forecast results of sub-season to season (S2S) for the next 1-6 weeks in the target area provided by the National Centers for Environmental Prediction (NCEP) of the United States, ERA5 historical meteorological data, digital elevation model (DEM) data, and Arctic polar vortex-related data;

[0007] Step 2: Preprocess the data. First, process the ERA5 historical meteorological data of the region to obtain daily average meteorological data, where the total precipitation element is statistically accumulated from hourly precipitation to daily precipitation, and when the total precipitation is used as a label, it is the cumulative precipitation from the forecast start date to the last day; the digital elevation model (DEM) data is processed through calculation to obtain altitude, slope, and aspect data; the Arctic polar vortex data is also preprocessed to obtain the geopotential height value for each day, and finally, all data is subjected to standardization calculation;

[0008] Step 3: Input the data related to the Digital Elevation Model (DEM) into the Multi-scale Feature Convolutional Fusion Network (MFCFN) to obtain the DEM features; input the historical meteorological data of ERA5 for the past 30 days into the Temporal-Spatial Attention Convolutional Long Short-Term Memory Network (T-SAConvLSTM) model to obtain the temporal-spatial features in the pre-forecast period; the polar vortex data is obtained through the teleconnection mapping method to get a mapping result with the same dimension as the forecast data, and then use the T-SAConvLSTM model network to further extract the temporal-spatial features; finally, obtain the feature maps of multi-source data.

[0009] Step 4: Input the multi-source data feature maps and the precipitation forecast data to be corrected into the Multi-source Cross Attention Fusion Module (MCAF) to obtain the weighted multi-source data feature maps.

[0010] Step 5: Input the weighted multi-source data feature maps into the correction module MDBU-Net that integrates the Max Pooling and Dense Connection Block (MDB) to obtain the corrected results of the sub-seasonal cumulative precipitation forecast for the next 1-6 weeks.

[0011] As a further preferred solution of the sub-seasonal precipitation forecast correction method based on MF-SPCM of the present invention, in Step 1, the precipitation forecast data, historical meteorological data, Digital Elevation Model (DEM) data, and Arctic polar vortex data are all grid data; among them, the spatial resolutions of the precipitation forecast data, historical meteorological data, and Arctic polar vortex data are all 0.25°, approximately 10 km, and the spatial resolution of the DEM data is 1 km; the precipitation forecast statistics are the cumulative precipitation, that is, the cumulative precipitation data from the forecast start date to the end date, and the time resolution of the historical meteorological data is 1 hour; the Arctic polar vortex data is the geopotential height data on the 10 hPa, 50 hPa, 70 hPa and other isobaric surfaces in the region of 60°N - 90°N, -180°E - 180°E, and the time resolution is 1 hour.

[0012] As a further preferred solution of the sub-seasonal precipitation forecast correction method based on MF-SPCM of the present invention, in step 3, the multi-scale feature convolutional fusion network MFCFN adopts a hierarchical architecture design, and realizes collaborative representation learning of multi-scale features through a multi-branch convolutional path and a cross-layer connection mechanism; specifically, the input three-channel DEM raw data includes altitude, slope, and aspect, which are processed by a deep feature extraction unit containing residual blocks to generate initial DEM features; the residual block contains two 3×3 convolutional layers, two Bn layers, and two ReLU activation function layers, and the residual connection branch contains a 1×1 convolutional layer and a Bn layer; then, in order to capture the multi-scale geometric features of the terrain data, three groups of heterogeneous convolutional kernels with sizes of 7×7, 5×5, and 3×3 are designed in the network, so that the model can extract terrain feature information at coarse-grained, medium-grained, and fine-grained spatial scales respectively; in the feature fusion stage, a two-stage strategy is adopted for feature reconstruction. Taking the layer with a convolutional kernel size of 5×5 as an example, the features of the previous layer are restored to the spatial resolution through an upsampling operation, and the features of this layer are reduced in the channel dimension by using a 1×1 convolutional layer; the semantic alignment of features at different levels is achieved through pixel-level addition, and the multi-scale feature maps are integrated through a channel splicing operation.

[0013] As a further preferred solution of the sub-seasonal precipitation forecast correction method based on MF-SPCM of the present invention, in step 3, the T-SAConvLSTM model network is implemented by a tandem architecture. The ERA5 historical meteorological data includes meteorological elements such as 2m temperature, 10m wind field, and humidity, and the time is 30 days of historical meteorological data from 30 days before the forecast start date to 1 day before the forecast start date; the first part of the TA module includes a global average pooling layer GAP, a channel-independent multi-layer perceptron CMLP, and a Softmax function; the meteorological data first extracts temporal features through GAP, then performs a non-linear transformation through CMLP, and then generates a time attention weight matrix with physical significance through the Softmax function. Finally, the model weight matrix is weighted with the input data to achieve adaptive enhancement of the key meteorological evolution stage; the second part is to embed spatial attention into the ConvLSTM unit to construct a SAConvLSTM network model. Finally, the two are jointly constructed into a physically interpretable spatio-temporal feature extractor, and the output data of the previous module is input into the SAConvLSTM model, and finally a historical meteorological element feature map is obtained; the calculation process is shown in equation (1): (1) where, represents element-wise multiplication, and represent the input data and the output data respectively.

[0014] As a further preferred solution of the sub-seasonal precipitation forecast correction method based on MF-SPCM of the present invention, in step 4, the multi-source cross-attention fusion module MCAF is composed of a self-attention module and a cross-attention module; first, the internal spatio-temporal correlations of the forecast data and multi-source data features are respectively modeled through a parallel self-attention mechanism. The forecast branch uses a self-attention layer to capture the potential cumulative relationships between forecast time periods, and the auxiliary feature branch enhances the spatial correlation representation of elements such as terrain and circulation through self-attention; subsequently, the two groups of feature tensors weighted by self-attention are input into the cross-attention layer, where the auxiliary features are used as query vectors and the forecast data is used as key-value pairs, and a cross-modal dynamic feature mapping relationship is established through a learnable attention weight matrix to realize the collaborative modeling of the evolution law of the meteorological system and multi-source meteorological factors; finally, the cross-attention output and the weighted features of the auxiliary feature branch are concatenated through the channel dimension and the dimension alignment and non-linear fusion of cross-modal features are realized through 1×1 convolution; the specific calculation process is as follows: (2) (3) (4) Among them, is the forecast query vector, is the forecast key vector, is the forecast value vector; is the key vector dimension, a normalization factor for scaling the attention score; is the normalization exponential function, which converts the calculation result into a probability distribution, maps the value to between 0 and 1 and the sum is 1; represents the forecast attention output matrix, the final feature representation obtained through weighted aggregation operation, which combines historical information and the focus of the current query; is the auxiliary feature query vector; is the auxiliary feature key vector; is the auxiliary feature value vector; represents 1×1 convolution, which is used to adjust the number of channels of the forecast data weighted features to be the same as that of the auxiliary data channels; is the auxiliary feature attention output matrix; 、 and are tensors respectively calculated by 、 ; Concat represents the channel concatenation operation; is the fusion attention matrix.

[0015] As a further preferred solution of the sub-seasonal precipitation forecast correction method based on MF-SPCM of the present invention, in step 3, the teleconnection mapping method includes three steps:

[0016] (1) Initialize an empty tensor matrix as the relevant field container based on the spatial dimension of the forecast data;

[0017] (2) Calculate the temporal correlation matrix between the concurrent ERA5 precipitation field and the polar vortex field grid by grid using the Pearson correlation coefficient algorithm, and identify the spatial coordinate points with the maximum correlation coefficient in the polar vortex field;

[0018] (3) Map the geopotential height value of the optimal correlation point to the corresponding ERA5 geospatial position in the empty matrix. The obtained mapping matrix is the mapping result of the teleconnection. The calculation process is shown in equations (5) to (8):

[0019] (5)

[0020] (6)

[0021] (7)

[0022] (8)

[0023] Among them, is the established empty tensor matrix, represents the empty tensor matrix, is the ERA5 precipitation field, is the polar vortex field, is the calculation result of the Pearson correlation coefficient, is the covariance, is the standard deviation; is the grid point position of the polar vortex field with the maximum correlation, is the value of the polar vortex field at this specific position; is the data carrier carrying the key information of the polar vortex field.

[0024] As a further preferred solution of the sub-seasonal precipitation prediction correction method based on MF-SPCM of the present invention, in step 5, the correction module MDBU-Ne performs spatial dimension compression on the input feature map using max pooling operation, suppressing noise interference while retaining the significant features of meteorological elements; subsequently, a convolutional block including a three-layer densely connected structure is constructed, where the input of each convolutional operation is composed of the output feature maps of all its previous layers along the channel dimension spliced together, that is, the number of input channels of the nth layer is the sum of the initial input channel number and the output channel number of the n-1th layer; the encoding stage deeply mines the spatio-temporal correlation features of the multi-source data fusion feature map through this cross-layer gradient propagation mechanism; finally, the coarse-grained features output by the max pooling path and the fine-grained features extracted by the dense convolution path are spliced along the channel dimension to form a composite feature expression with both spatial robustness and detail resolution ability; in the decoding stage, when performing upsampling operation through transposed convolution, first perform transposed convolution operation on the low-resolution feature map to restore the spatial resolution of the feature map to the size of the corresponding level of the encoder; then obtain the multi-scale features output by the same layer of the encoder through skip connection, and the two achieve cross-level feature fusion through channel splicing; finally, the model obtains the cumulative precipitation results for the next 1-6 weeks.

[0025] The present invention adopts the above technical solutions and has the following technical effects compared with the prior art:

[0026] (1) In terms of data, the present invention considers the influence mechanism of multi-source data and incorporates data of polar vortices as an aid, improving the accuracy of prediction correction.

[0027] (2) The present invention adopts channel and spatio-temporal attention mechanisms to strengthen the weights of key meteorological factors, time, and space, ensuring the effective utilization of historical meteorological data to the greatest extent.

[0028] (3) The present invention adopts a multi-scale feature extraction model, and different-sized convolutional kernels are used in the convolutional layer for operations, expanding the receptive field to improve the generalization ability of the model and enhancing the robustness and adaptability of the model.

[0029] (4) When the present invention incorporates data of polar vortices, the grid data with the strongest correlation is selected. On the one hand, it reduces the data dimension, which is beneficial to the calculation of the end-to-end model; on the other hand, it reduces the influence of noise data and improves the calculation efficiency of the deep learning model. Description of the Drawings

[0030] Figure 1 It is a flowchart of a sub-seasonal precipitation prediction correction method based on MF-SPCM proposed by the present invention;

[0031] Figure 2 It is a model diagram of a sub-seasonal precipitation prediction correction method based on MF-SPCM proposed by the present invention;

[0032] Figure 3 This is the TA model structure diagram of the present invention;

[0033] Figure 4 This is the structural schematic diagram of the spatial attention SA;

[0034] Figure 5 This is the structural schematic diagram of the SAConvLSTM model;

[0035] Figure 6 This is the structural schematic diagram of the MFCFN model;

[0036] Figure 7 This is the structural schematic diagram of the MCAF model;

[0037] Figure 8 This is the structural schematic diagram of the MDB module;

[0038] Figure 9 This is the result diagram of the ablation experiment;

[0039] Figure 10 This is the result diagram of the comparative experiment. Detailed implementation manner

[0040] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings:

[0041] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention. The present invention will be described in detail below according to the accompanying drawings and preferred embodiments, and the purpose and effect of the present invention will become more apparent. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0042] The technical solution of the present invention will be further elaborated below in combination with the accompanying drawings and embodiments.

[0043] Embodiment 1:

[0044] As Figure 1 and Figure 2 shown, this embodiment proposes a sub-seasonal precipitation forecast correction method based on MF-SPCM, which mainly includes the following steps:

[0045] Step 1: Obtain the sub-seasonal precipitation forecast results, ERA5 historical meteorological data, digital elevation model (DEM) data, and Arctic polar vortex-related data for the target area in the next 1 - 6 weeks provided by NCEP (National Centers for Environmental Prediction). The precipitation forecast data, historical meteorological data, DEM data, and polar vortex data are all grid data. Among them, the spatial resolutions of the precipitation forecast, historical meteorological, and polar vortex data are all 0.25°, approximately 10 km, and the spatial resolution of the DEM data is 1 km. The precipitation forecast statistics are cumulative precipitation, that is, the cumulative precipitation data from the forecast start date to the end date. The time resolution of the historical meteorological data is 1 hour. The polar vortex data is the geopotential height data on the 10 hPa, 50 hPa, and 70 hPa isobaric surfaces in the region of 60°N - 90°N, -180°E - 180°E, and the time resolution is 1 hour. The specific operations include:

[0046] In this embodiment, the middle and lower reaches of the Yangtze River are located in the range of 24.5° - 32°N, 110° - 122°E. Obtain the meteorological historical actual data of this area from 2015 to 2024 from the ERA5 dataset released by the European Centre for Medium-Range Weather Forecasts (ECMWF). The time resolution is 1 hour, and the spatial resolution is 0.25°. The meteorological elements include 12 meteorological elements such as total precipitation, temperature, dew point temperature, wind speed, humidity, and surface pressure. The DEM data is from the Geospatial Cloud Platform. The data spatial resolution is 30 m, and after resampling, it is 1 km resolution, and then the slope and aspect data at 1 km resolution are calculated. The polar vortex data is from the ERA5 dataset, and the geopotential height data on the three isobaric surfaces of 10 hPa, 50 hPa, and 70 hPa in the region of 60°N - 90°N, -180°E - 180°E is used.

[0047] Step 2: Preprocess the data. First, process the ERA5 historical meteorological data of the region to obtain the daily average meteorological data. Among them, the total precipitation element is statistically calculated from the hourly precipitation to the daily cumulative precipitation. When the total precipitation is used as a label, it is the cumulative precipitation from the forecast start date to the last day. The DEM data is processed through calculation to obtain elevation, slope, and aspect data. The polar vortex data also obtains the daily geopotential height value through preprocessing. Finally, perform standardization calculations on all data. The specific operations include:

[0048] In this embodiment, first, the daily mean values of the other 11 meteorological element data are calculated, and for the total precipitation, the daily cumulative precipitation is obtained. When the total precipitation is used as a label, the cumulative precipitation from the forecast start date to the last date is statistically calculated; the polar vortex data is also calculated as the daily mean value; the formula used for the standardization method is:

[0049] (1)

[0050] In the formula, is the input data, is the data after standardization, is the mean value of the input data, is the standard deviation of the input data.

[0051] Step 3: Input the data related to the digital elevation model DEM into the multi-scale feature convolutional fusion network MFCFN to obtain the digital elevation model DEM features; input the historical meteorological data of ERA5 in the past 30 days into the temporal-spatial attention convolutional long short-term memory network T-SAConvLSTM (Temporal-Spatial Attention Convolutional LSTM) model to obtain the temporal-spatial features in the pre-forecast period; the polar vortex data is obtained through the teleconnection mapping method to get a mapping result with the same dimension as the forecast data, and then the T-SAConvLSTM model network is used to further extract the temporal-spatial features; finally, the feature map of multi-source data is obtained. The specific operations include:

[0052] In this embodiment, the MFCFN network is as Figure 6As shown in the figure, the workflow is as follows: First, the standardized elevation, slope, and aspect data are input into the residual block to obtain the initial terrain features. The residual block contains two 3×3 convolutional layers, two Bn layers, and two ReLU activation function layers. Additionally, the residual connection branch contains a 1×1 convolution and a Bn layer. Then, in order to capture terrain features at multiple scales, the initial features are successively passed through three groups of heterogeneous convolutional layers with kernel sizes of 7×7, 5×5, and 3×3 respectively. This design enables the model to extract terrain feature information at coarse-grained, medium-grained, and fine-grained spatial scales respectively. In the feature fusion stage, a two-stage strategy is adopted for feature reconstruction. Taking the layer with a 5×5 convolutional kernel as an example, first, the features of the previous layer are restored to the spatial resolution through upsampling operations, and at the same time, the features of this layer are reduced in the channel dimension using 1×1 convolution. On this basis, pixel-level addition is used to achieve semantic alignment of features at different levels, and multi-scale feature maps are integrated through channel concatenation operations. To enhance the local feature representation ability and the adjustment of spatial resolution, a series of normalization operation modules are integrated in each convolutional processing layer of this network, specifically including batch normalization, convolution operations, average pooling, and ReLU non-linear activation functions. This modular design not only maintains the balance between network depth and computational efficiency, but also the multi-scale feature cross-fusion mechanism helps to improve the extraction accuracy of fine-grained features such as terrain edges and textures, providing a robust terrain feature representation basis for subsequent improvement of precipitation forecasting tasks. Finally, the MFCFN model outputs a terrain feature map with a dimension of (3, 31, 49);

[0053] The described TA model, SA model, and SAConvLSTM network structures are respectively as Figure 3 , Figure 4 and Figure 5As shown in the figure. The core algorithm of the TA model realizes dynamic time weight reconstruction through multi-stage feature transformation. First, a channel-oriented global average pooling operation is performed on the input ERA5 meteorological data. In view of the heterogeneity characteristics of the temporal influence of meteorological elements, this model independently performs GAP calculations for 12 meteorological elements respectively, reducing the original four-dimensional tensor of 12×30×31×49 to a time weight initial value matrix of 12×30×1×1. This process extracts the initial significant features of each meteorological element at different time steps through the mean operation along the spatial dimensions (H, W). However, as a static calculation method, traditional GAP can only capture linear temporal correlations, but it is difficult to represent the non-linear interaction effects of multi-scale processes in the atmospheric system. To make up for the limitations of GAP, the present invention constructs a CMLP model after the GAP layer. This model adopts a hierarchical architecture of "grouping - dimension elevation - non-linear activation - dimension reduction", which realizes non-linear mapping of the temporal feature space and mining of high-order interaction relationships while ensuring computational efficiency. In addition, it also controls that each meteorological element is not interfered by the time weights of other elements. In addition, this design ensures that the time weights of each meteorological element are calculated independently, avoiding interference between different elements, so as to balance the feature expression ability and the model generalization ability. The 12×30×1×1 feature tensor obtained after CMLP transformation is finally normalized along the time dimension through the Softmax function to generate dynamic time attention weights, so as to accurately quantify the contribution degree differences of meteorological elements at different historical moments to the predicted precipitation. The calculation formula is as follows:

[0054] (2)

[0055] where, represents element-wise multiplication, and represent the input data and the output data respectively.

[0056] The SA-ConvLSTM model optimizes the spatio-temporal feature extraction ability of the traditional ConvLSTM by embedding a spatial attention mechanism. Its core structure is as Figure 5 shown. This model integrates a spatial attention mechanism at the input end of ConvLSTM, and inputs the meteorological element input data at the current moment and the hidden state at the previous moment into the spatial attention model for calculation after channel concatenation. Inside the spatial attention model SA, it includes max pooling, average pooling and the Softmax function. The structure is as Figure 4As shown. Through these calculation processes, a feature map with a spatial weight distribution is generated. This design enables the model to dynamically calculate the importance weights of different meteorological elements at each grid point position, achieve precise feature weighting for key spatial regions, and effectively enhance the spatial feature capture ability of ConvLSTM in complex meteorological fields. Finally, the model reduces the dimensionality of the spatio-temporal features of the 12 meteorological elements extracted, and finally outputs an ERA5 historical meteorological element feature map of 12×31×49.

[0057] The calculation process of ConvLSTM is shown in Equations (3) to (7).

[0058] (3)

[0059] (4)

[0060] (5)

[0061] (6)

[0062] (7)

[0063] Among them, represents the spatio-temporal meteorological field input at time t. The input gate , the forget gate , and the output gate are dynamically generated gate coefficients in the range of 0 to 1 through convolution operations and the Sigmoid activation function. Among them, the input gate is used to update the cell state, the forget gate is used to select the meteorological information input at time t for forgetting, and the output gate is to control the spatio-temporal features of the current output. These gating mechanisms are all calculated according to the input at the current time t and the cell state at the previous time t - 1 to obtain the corresponding gate coefficients, as shown in Equations 5 to 7, where represents the convolution operation, represents the Sigmoid operation, is the convolution kernel weight, is the bias term. The cell state is jointly regulated by the forget gate and the input gate to achieve selective memory of historical information, as shown in Equation 8, where is the element-wise multiplication, is the long-term memory of the previous time; the output gate regulates the information output of the cell state to generate the current hidden state , where tanh represents the tanh activation function.

[0064] The technical process of the teleconnection mapping method can be decomposed into three core steps: First, initialize a null-value tensor matrix as the correlation field container based on the spatial dimension of the forecast data. The dimension of this null-value tensor matrix is (30, 3, 31, 49). Second, calculate the temporal correlation matrix between the concurrent ERA5 precipitation field and the polar vortex field grid by grid through the Pearson correlation coefficient algorithm, and identify the spatial coordinate points with the maximum correlation coefficient in the polar vortex field. Finally, map the geopotential height value of this optimal correlation point to the corresponding ERA5 geographical space position in the null matrix. The obtained mapping matrix is the mapping result of the teleconnection. The calculation process is shown in Equations (8) to (11).

[0065] (8)

[0066] (9)

[0067] (10)

[0068] (11)

[0069] Among them, is the established null tensor matrix, represents the null tensor matrix, is the ERA5 precipitation field, is the polar vortex field. is the calculation result of the Pearson correlation coefficient, is the covariance, is the standard deviation; is the grid position of the polar vortex field with the maximum correlation, is the value of the polar vortex field at this specific position; is the data carrier carrying the key information of the polar vortex field.

[0070] After teleconnection mapping, the polar vortex data changes from the original dimension of (30, 3, 121, 1440) to grid data with a stronger correlation with precipitation, and its dimension is (30, 3, 31, 49). Then, the spatio-temporal features of this tensor are further extracted through the above T-SAConvLSTM (Temporal-Spatial Attention Convolutional LSTM) network model. Finally, the dimension of the obtained spatio-temporal feature map is (3, 31, 49);

[0071] Step 4: Input the above multi-source data feature map and the precipitation forecast data to be corrected into the multi-source cross-attention fusion module (Multi-source Cross-Attention Fusion, MCAF) to obtain the weighted multi-source data feature map;

[0072] The MCAF (Multi-source Cross-Attention Fusion) module is composed of a self-attention module and a cross-attention module. Its structure is as follows: Figure 7 As shown. First, the internal spatiotemporal correlation of forecast data and multi-source data features are modeled separately through the parallel self-attention mechanism. The forecast branch uses the self-attention layer to capture the potential cumulative relationship between each forecast period, and the auxiliary feature branch uses self-attention to enhance the spatial correlation representation of factors such as terrain and circulation. Subsequently, the two sets of feature tensors weighted by self-attention are input into the cross-attention layer, where the auxiliary features are used as query vectors and the forecast data are used as key-value pairs. A cross-modal dynamic feature mapping relationship is established through a learnable attention weight matrix to achieve collaborative modeling of the evolution law of the meteorological system and multi-source meteorological factors. Finally, the cross-attention output and the weighted features of the auxiliary feature branch are spliced ​​through the channel dimension, and the dimensional alignment and nonlinear fusion of cross-modal features are achieved through 1×1 convolution. The specific calculation process is as follows:

[0073] (12)

[0074] (13)

[0075] (14)

[0076] in, is the forecast query vector, is the prediction key vector, is the forecast value vector; is the key vector dimension, a normalization factor used to scale the attention score; To normalize the exponential function, convert the calculation results into probability distribution, map the values ​​between 0 and 1 and the sum is 1; Represents the predicted attention output matrix, the final feature representation obtained through the weighted aggregation operation, which combines historical information and the focus of the current query; is the auxiliary feature query vector; is the auxiliary feature key vector; is the auxiliary eigenvalue vector; represents 1×1 convolution, which is used to adjust the number of channels of the weighted features of the forecast data to make it consistent with the number of channels of the auxiliary data; Output matrix for auxiliary feature attention; , and They are respectively , The calculated tensor; Concat represents the channel concatenation operation; is the fusion attention matrix.

[0077] Step 5: Input the weighted multi-source data feature map into the MDBU-Net correction module that integrates the Maxpool and Dense Block (MDB), and finally obtain the corrected results of the sub-seasonal cumulative precipitation forecast for the next 1-6 weeks.

[0078] The structure of the MDB module is as Figure 8 shown. This module uses the max-pooling operation to perform spatial dimension compression on the input feature map (downsampling to 1 / 4 of the original size), suppressing noise interference while retaining the significant features of meteorological elements; subsequently, a convolutional block containing a three-layer dense connection structure is constructed, where the input of each convolutional operation is composed of the output feature maps of all its previous layers concatenated along the channel dimension, that is, the number of input channels of the nth layer is the sum of the initial input channels and the number of output channels of the n-1 layer. This dense feed-forward characteristic significantly enhances the gradient flow efficiency and feature expression ability by establishing a cross-layer feature reuse mechanism. In the encoding stage of MDBU-Net, it is precisely through this cross-layer gradient propagation mechanism that the spatio-temporal correlation features of the multi-source data fusion feature map are deeply mined; finally, the coarse-grained features output by the max-pooling path and the fine-grained features extracted by the dense convolutional path are concatenated along the channel dimension to form a composite feature expression with both spatial robustness and detail resolution ability. This two-stream fusion architecture effectively alleviates the problem of feature resolution attenuation caused by the traditional downsampling process through the collaborative mechanism of feature selection and enhancement while ensuring computational efficiency.

[0079] In the decoding stage, when performing upsampling through transposed convolution, first perform a transposed convolution operation on the low-resolution feature map to restore the spatial resolution of the feature map to the size of the corresponding layer of the encoder; then obtain the multi-scale features output by the same layer of the encoder through skip connections, and the two are fused through channel concatenation to achieve cross-level feature fusion. Finally, the model obtains the cumulative precipitation results for the next 1-6 weeks.

[0080] Use the relevant data from May to August from 2015 to 2023 as the training set, and the training label is the cumulative precipitation value for the next 1 week to 6 weeks. Use the Adam optimizer to train the model, with a learning rate of 0.001, and the loss function uses the mean absolute error MAE. The formula is:

[0081] (15)

[0082] where is the number of samples, is the true value of the th sample, is the corrected forecast value of the th sample.

[0083] Use the trained model to predict the test set data to judge the accuracy of the model.

[0084] In this example, ablation experiments were set up to verify the effectiveness of each module, and comparative experiments were conducted to prove the accuracy advantage of the proposed model compared with existing advanced methods. During the experiment, the root mean square error (RMSE) was used as the evaluation index, and the calculation formula is as follows:

[0085] (16)

[0086] where represents the number of test samples, is the number of grid points in the correction area, represents the historical true value of ERA5, represents the output result of the correction model, is the th sample in the test sample, is the th grid point of the grid data in the correction area.

[0087] First, ablation experiments were conducted. The results of the ablation experiments from May to August 2024 are shown in the table, and the quantitative results are shown in Table 1. The visualization results of the RMSE distribution are as Figure 9 shown. It can be clearly seen from Table 1 that the RMSE error of the MF-SPCM network model correction method of the present invention is the smallest, and the RMSE error gradually decreases when gradually integrating multi-source data modules, indicating that these modules play a positive role. As Figure 9 shown, MF-SPCM exhibits the best error performance. With the gradual introduction of different data, the RMSE index of the model shows a significant downward trend. This phenomenon not only verifies the improvement effect of incremental data on the model performance but also highlights the effectiveness of data quality and model design.

[0088] Table 1

[0089] Method MFCFN T-SAConvLSTM Teleconnection mapping MCAF MDB RMSE NCEP forecast result 123.9864 DEM ✔ 86.3633 DEM-ERA5 ✔ ✔ 85.0876 Multi-source data ✔ ✔ ✔ 83.2892 MCAF ✔ ✔ ✔ ✔ 82.5379 MF-SPCM ✔ ✔ ✔ ✔ ✔ 81.6757

[0090] In the comparative experiment, the NCEP forecast results were compared with the correction results of multiple advanced models, namely A-ConvLSTM, PDFN, CSG-UNET, and EFC-Net. The results of the comparative experiment from May to August 2024 are shown in the table, and the quantitative results are shown in Table 2. The visualization results are as Figure 10 shown. It can be seen from the figure and table results that compared with these advanced models, the MF-SPCM proposed by the present invention has obvious advantages and shows the lowest error value of the correction result.

[0091] Table 2

[0092] Index NCEP A-ConvLSTM PDFN CSG-UNET EFC-Net MF-SPCM RMSE 123.9864 93.5689 84.6914 94.3823 87.0681 81.6757

[0093] A deep learning method for sub-seasonal precipitation forecast correction based on multi-source data fusion proposed by the present invention includes topographic data and meteorological element data of the region where it is located in terms of data, and innovatively fuses polar vortex data; in terms of methods, different neural network modules are used to process multi-source data, and at the same time, the characteristic information beneficial to precipitation forecast contained in these data is extracted, so as to improve the accuracy of precipitation forecast; in the backbone network of the end-to-end model, an improved MDBU-Net is used to perform in-depth complex calculations. This model adopts the construction idea of residuals to reduce the information loss during the operation of the network.

[0094] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above various methods. Among them, any reference to a memory, storage, database or other medium used in the various embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0095] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combinations of these technical features do not conflict, they should be considered as the scope described in this specification.

[0096] The above-described embodiments merely represent several implementation manners of the present application. Their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A sub-seasonal precipitation forecast correction method based on MF-SPCM, characterized in that: Specifically, it includes the following steps: Step 1: Obtain the sub-seasonal precipitation forecast results, ERA5 historical meteorological data, digital elevation model (DEM) data, and Arctic polar vortex-related data for the next 1 - 6 weeks in the target area provided by the National Centers for Environmental Prediction (NCEP) of the United States; Step 2: Preprocess the data. First, process the ERA5 historical meteorological data of the region to obtain daily mean meteorological data. Among them, the total precipitation element is statistically accumulated from hourly precipitation to daily cumulative precipitation. When the total precipitation is used as a label, it is the cumulative precipitation from the forecast start date to the last day; the digital elevation model (DEM) data is processed through calculation to obtain elevation, slope, and aspect data; the Arctic polar vortex data also obtains the geopotential height value of each day through preprocessing. Finally, standardize all the data; Step 3: Input the digital elevation model (DEM) - related data into the multi - scale feature convolutional fusion network (MFCFN) to obtain the digital elevation model (DEM) features; Input the ERA5 historical meteorological data of the past 30 days into the spatio - temporal attention convolutional long - short - term memory network (T - SAConvLSTM) model to obtain the spatio - temporal features in the pre - forecast period; the polar vortex data obtains a mapping result with the same dimension as the forecast data through the teleconnection mapping method, and then uses the T - SAConvLSTM model network to further extract the spatio - temporal features; finally, obtain the feature map of multi - source data; Step 4: Input the multi - source data feature map and the precipitation forecast data to be corrected into the multi - source cross - attention fusion module (MCAF) to obtain the weighted multi - source data feature map; Step 5: Input the weighted multi - source data feature map into the correction module (MDBU - Net) that integrates the max - pooling and dense connection block (MDB) to obtain the corrected results of the sub - seasonal cumulative precipitation forecast for the next 1 - 6 weeks.

2. A sub-seasonal precipitation forecast correction method based on MF-SPCM according to claim 1, characterized in that: In Step 1, the precipitation forecast data, historical meteorological data, digital elevation model (DEM) data, and Arctic polar vortex data are all grid data; among them, the spatial resolutions of the precipitation forecast data, historical meteorological data, and Arctic polar vortex data are all 0.25°, 10 km, and the spatial resolution of the DEM data is 1 km; the precipitation forecast statistics the cumulative precipitation, that is, the cumulative precipitation data from the forecast start date to the end date, and the time resolution of the historical meteorological data is 1 hour; the Arctic polar vortex data is the geopotential height data on the 10 hPa, 50 hPa, 70 hPa and other isobaric surfaces in the region of 60°N - 90°N, - 180°E - 180°E, and the time resolution is 1 hour.

3. A sub-seasonal precipitation forecast correction method based on MF-SPCM according to claim 1, characterized in that: In Step 3, the multi - scale feature convolutional fusion network (MFCFN) adopts a hierarchical architecture design and realizes the collaborative representation learning of multi - scale features through multi - branch convolutional paths and cross - layer connection mechanisms; specifically, the input three - channel DEM raw data includes elevation, slope, and aspect, which are processed by a deep feature extraction unit containing residual blocks to generate initial DEM features; The residual block contains two 3×3 convolutional layers, two Bn layers, and two ReLU activation function layers. Additionally, the residual connection branch contains a 1×1 convolutional layer and a Bn layer. Then, to capture the multi-scale geometric features of the terrain data, three groups of heterogeneous convolutional kernels with sizes of 7×7, 5×5, and 3×3 are designed in the network, enabling the model to extract terrain feature information at coarse-grained, medium-grained, and fine-grained spatial scales respectively. In the feature fusion stage, a two-stage strategy is adopted for feature reconstruction. Taking the layer with a 5×5 convolutional kernel as an example, the features of the previous layer are restored to the spatial resolution through upsampling operations, while the features of this layer are reduced in the channel dimension using 1×1 convolutions. Pixel-level addition is used to achieve semantic alignment of features at different levels, and multi-scale feature maps are integrated through channel concatenation operations.

4. A sub-seasonal precipitation forecast correction method based on MF-SPCM according to claim 1, characterized in that, In step 3, the T-SAConvLSTM model network is implemented by a tandem architecture. The ERA5 historical meteorological data includes meteorological elements such as 2m temperature, 10m wind field, and humidity, and the time period is 30 days of historical meteorological data from 30 days before the forecast start date to 1 day before the forecast start date. The TA module in the first part includes a global average pooling layer GAP, a channel-independent multi-layer perceptron CMLP, and a Softmax function. The meteorological data first extracts temporal features through GAP, then undergoes a non-linear transformation via CMLP, and then generates a time attention weight matrix with physical significance through the Softmax function. Finally, the model weight matrix is weighted with the input data to achieve adaptive enhancement of key meteorological evolution stages. The second part embeds spatial attention into the ConvLSTM unit to construct the SAConvLSTM network model. Finally, the two are jointly constructed into a physically interpretable spatio-temporal feature extractor. The output data of the previous module is input into the SAConvLSTM model, and finally, a historical meteorological element feature map is obtained. The calculation process is shown in Equation (1): (1) Among them, represents element-wise multiplication, and represent the input data and the output data respectively.

5. A sub-seasonal precipitation forecast correction method based on MF-SPCM according to claim 1, characterized in that, In step 4, the multi-source cross-attention fusion module MCAF consists of a self-attention module and a cross-attention module. First, the internal spatio-temporal correlations of the forecast data and multi-source data features are respectively modeled through a parallel self-attention mechanism. The forecast branch uses the self-attention layer to capture the potential cumulative relationships between different forecast times, and the auxiliary feature branch enhances the spatial correlation representation of terrain and circulation elements through self-attention. Subsequently, the two groups of feature tensors weighted by self-attention are input into the cross-attention layer, where the auxiliary features serve as query vectors and the forecast data serves as key-value pairs. A cross-modal dynamic feature mapping relationship is established through a learnable attention weight matrix to achieve collaborative modeling of the evolution law of the meteorological system and multi-source meteorological factors. Finally, the output of the cross-attention and the weighted features of the auxiliary feature branch are concatenated in the channel dimension, and 1×1 convolutions are used to achieve dimension alignment and non-linear fusion of cross-modal features. The specific calculation process is as follows: (2) (3) (4) Among them, is the forecast query vector, is the forecast key vector, is the forecast value vector; is the key vector dimension, a normalization factor for scaling the attention score; is the normalization exponential function, which converts the calculation result into a probability distribution, maps the values to between 0 and 1 and the sum is 1; represents the forecast attention output matrix, the final feature representation obtained through weighted aggregation operation, combining historical information and the focus of the current query; is the auxiliary feature query vector; is the auxiliary feature key vector; is the auxiliary feature value vector; represents a 1×1 convolution, which is used to adjust the number of channels of the weighted features of the forecast data to be the same as the number of channels of the auxiliary data; is the auxiliary feature attention output matrix; , and are tensors calculated from , respectively; Concat represents the channel concatenation operation; is the fused attention matrix.

6. A sub-seasonal precipitation forecast correction method based on MF-SPCM according to claim 1, characterized in that: In step 3, the teleconnection mapping method includes three steps: (1) Initialize an empty tensor matrix as the correlation field container based on the spatial dimension of the forecast data; (2) Calculate the temporal correlation matrix between the concurrent ERA5 precipitation field and the polar vortex field at each grid point through the Pearson correlation coefficient algorithm, and identify the spatial coordinate points with the maximum correlation coefficient in the polar vortex field; (3) Map the geopotential height values of the optimal correlation points to the corresponding ERA5 geographical spatial positions in the empty matrix, and the obtained mapping matrix is the mapping result of the teleconnection. The calculation process is shown in Equations (5) to (8): (5) (6) (7) (8) Among them, is the established empty tensor matrix, represents the empty tensor matrix, is the ERA5 precipitation field, is the polar vortex field, is the calculation result of the Pearson correlation coefficient, is the covariance, is the standard deviation; is the grid point position of the polar vortex field with the maximum correlation, is the value of the polar vortex field at this specific position; is the data carrier carrying the key information of the polar vortex field.

7. A sub-seasonal precipitation forecast correction method based on MF-SPCM according to claim 1, characterized in that: In step 5, the correction module MDBU-Ne performs spatial dimension compression on the input feature map using max pooling operation, suppressing noise interference while retaining the significant features of meteorological elements; subsequently, a convolutional block containing a three-layer densely connected structure is constructed, where the input of each convolutional operation is composed of the output feature maps of all its previous layers concatenated along the channel dimension, that is, the number of input channels of the nth layer is the sum of the initial input channel number and the output channel number of the n - 1th layer; it is through this cross-layer gradient propagation mechanism in the encoding stage that the spatio-temporal correlation features of the multi-source data fusion feature map are deeply mined; finally, the coarse-grained features output by the max pooling path and the fine-grained features extracted by the dense convolution path are concatenated along the channel dimension to form a composite feature expression with both spatial robustness and detail resolution ability; In the decoding stage, when performing upsampling through transposed convolution, first perform transposed convolution operation on the low-resolution feature map to restore the spatial resolution of the feature map to the size of the corresponding layer of the encoder; subsequently, obtain the multi-scale features output by the same layer of the encoder through skip connection, and the two achieve cross-level feature fusion through channel concatenation; finally, the model obtains the cumulative precipitation results for the next 1 - 6 weeks.

Images