A correction method for subseasonal precipitation forecast based on MF-SPCM

Abstract

The present invention discloses a sub-seasonal precipitation forecast correction method based on MF-SPCM, which relates to the technical field of precipitation forecast correction. The method comprises the following steps: obtaining S2S sub-seasonal precipitation forecast results for a target area for the next 1-6 weeks provided by NCEP, ERA5 historical meteorological data, digital elevation model (DEM) data, and Arctic polar vortex-related data; pre-processing the data to obtain digital elevation model (DEM) features, spatiotemporal features before the forecast, and mapping results consistent with the dimensions of the forecast data, and extracting the spatiotemporal features; obtaining a feature map of multi-source data; inputting the multi-source data feature map and the precipitation forecast data to be corrected into a multi-source cross-attention fusion module to obtain a weighted multi-source data feature map; obtaining a corrected sub-seasonal cumulative precipitation forecast result for the next 1-6 weeks; and effectively improving the accuracy and timeliness of the sub-seasonal precipitation forecast.

Description

Technical Field

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

[0002] Subseasonal meteorological forecasts, which fall between weather forecasts and seasonal forecasts, have received relatively little attention and are often referred to as "forecast deserts." They also provide valuable guidance and insights for industries closely related to weather and climate conditions, such as power dispatch, fisheries, and agricultural development. However, current subseasonal precipitation forecasts still exhibit significant bias, making them insufficient for practical forecasting. Therefore, there is an urgent need to develop methods to improve the accuracy of precipitation forecasts on a subseasonal 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 to address the shortcomings of the background technology, so as to effectively improve the accuracy and timeliness of sub-seasonal precipitation forecast.

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

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

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

[0007] Step 2: Data preprocessing: First, the regional ERA5 historical meteorological data is processed to obtain daily mean meteorological data. The total precipitation element is calculated from hourly precipitation statistics to daily cumulative precipitation. When total precipitation is used as a label, it is the cumulative precipitation from the forecast start day to the last day. Digital elevation model (DEM) data is calculated and processed to obtain elevation, slope, and aspect data. Arctic polar vortex data is also preprocessed to obtain daily potential height values. Finally, all data are standardized.

[0008] Step 3: Input the DEM (Digital Elevation Model) data into the Multi-Scale Feature Convolution Fusion Network (MFCFN) to obtain DEM features. Input the ERA5 historical meteorological data from the past 30 days into the Spatiotemporal Attention Convolutional Long Short-Term Memory (T-SAConvLSTM) model to obtain the spatiotemporal features of the early forecast period. The polar vortex data is mapped using teleconnection mapping to obtain a mapping result consistent with the forecast data dimension. The T-SAConvLSTM model network is then used to further extract spatiotemporal features. Finally, a feature map of the multi-source data is obtained.

[0009] 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;

[0010] Step 5: Input the weighted multi-source data feature map into the correction module MDBU-Net, which integrates the maximum pooling and dense connection block MDB, to obtain the revised results of the sub-seasonal cumulative precipitation forecast for the next 1-6 weeks.

[0011] As a further preferred embodiment of the MF-SPCM-based sub-seasonal precipitation forecast correction method 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; the spatial resolution of the precipitation forecast data, historical meteorological data, and Arctic polar vortex data is 0.25°, approximately 10 km, and the spatial resolution of the DEM data is 1 km; the precipitation forecast statistics are cumulative precipitation, i.e., 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 are geopotential height data on 10 hPa, 50 hPa, and 70 hPa isopotential surfaces in the 60°N-90°N, -180°E-180°E region, with a time resolution of 1 hour.

[0012] As a further preferred embodiment of the MF-SPCM-based sub-seasonal precipitation forecast correction method of the present invention, in step 3, the multi-scale feature convolution fusion network MFCFN adopts a hierarchical architecture design, and realizes the collaborative representation learning of multi-scale features through a multi-branch convolution 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 a residual block to generate initial DEM features; wherein the residual block contains two 3×3 convolution layers, two Bn layers, and two ReLU activation function layers, and the residual connection branch contains a 1×1 convolution layer and A Bn layer; then, in order to capture the multi-scale geometric features of terrain data, three sets of heterogeneous convolution kernels are designed in the network, with sizes of 7×7, 5×5, and 3×3, respectively, 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 convolution kernel size of 5×5 as an example, the spatial resolution of the previous layer is restored by upsampling operation, and the channel dimension of the features of this layer is reduced by 1×1 convolution; the semantic alignment of features at different levels is achieved through pixel-level addition, and the multi-scale feature maps are integrated through 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 serial architecture, and the ERA5 historical meteorological data includes meteorological elements such as 2m temperature, 10m wind field, and humidity. The time is from 30 days before the forecast start date to 1 day before the forecast start date, a total of 30 days of historical meteorological data; wherein the TA module of the first part includes a global average pooling layer GAP, a channel-independent multi-layer perceptron CMLP, and a Softmax function; the meteorological data is first extracted with time series features by GAP, and then nonlinearly transformed by CMLP, and then a time attention weight matrix with physical significance is generated by the Softmax function, and 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, and finally the two are combined to construct a spatiotemporal feature extractor with physical interpretability, and the output data of the previous module is input into the SAConvLSTM model, and finally the historical meteorological element feature map is obtained; the calculation process is shown in formula (1):

[0014] (1)

[0015] in, stands for element-wise multiplication, and Represents input data and output data respectively.

[0016] As a further preferred embodiment of the MF-SPCM-based subseasonal precipitation forecast correction method 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 spatiotemporal correlation of the forecast data and multi-source data features is modeled separately through a 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, 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:

[0017] (2)

[0018] (3)

[0019] (4)

[0020] in, is the forecast query vector, is the prediction key vector, is the forecast value vector; is the key vector dimension, which is the 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 sum to 1; Represents the predicted attention output matrix, the final feature representation obtained by 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 splicing operation; is the fusion attention matrix.

[0021] As a further preferred embodiment of the MF-SPCM-based sub-seasonal precipitation forecast correction method of the present invention, in step 3, the teleconnection mapping method includes three steps:

[0022] (1) Initialize a null value tensor matrix as the correlation field container based on the spatial dimension of the forecast data;

[0023] (2) The temporal correlation matrix between the ERA5 precipitation field and the polar vortex field during the same period was calculated grid by grid using the Pearson correlation coefficient algorithm, and the spatial coordinate point with the maximum correlation coefficient in the polar vortex field was identified;

[0024] (3) The potential height value of the optimal correlation point is mapped to the corresponding ERA5 geographic space position in the empty matrix. The resulting mapping matrix is the mapping result of the teleconnection. The calculation process is shown in Equations (5) to (8):

[0025] (5)

[0026] (6)

[0027] (7)

[0028] (8)

[0029] in, is the empty tensor matrix created, represents an empty tensor matrix, For the ERA5 precipitation field, is the polar vortex field, is the result of Pearson correlation coefficient calculation, is the covariance, is the standard deviation; is the polar vortex field grid point position with maximum correlation, is the value of the polar vortex field at that particular location; It is a data carrier that carries key information of the polar vortex field.

[0030] As a further preferred embodiment of the present invention's method for correcting subseasonal precipitation forecasts based on MF-SPCM, in step 5, the correction module MDBU-Ne uses a maximum pooling operation to perform spatial dimension compression on the input feature map, thereby suppressing noise interference while retaining the significant features of meteorological elements; a convolution block with a three-layer dense connection structure is then constructed, wherein the input of each convolution operation is composed of the output feature maps of all preceding layers spliced together along the channel dimension, i.e., the number of input channels of the nth layer is the sum of the number of initial input channels and the number of output channels of the n-1th layer; the encoding stage is precisely through this cross-layer gradient propagation mechanism to deeply mine the feature maps. The spatiotemporal correlation features of the multi-source data fusion feature map are explored; finally, the coarse-grained features output by the maximum pooling path and the fine-grained features extracted by the dense convolution path are spliced in the channel dimension to form a composite feature expression with both spatial robustness and detail resolution capabilities; in the decoding stage, when upsampling operation is achieved through deconvolution, deconvolution operation is first performed 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, the multi-scale features of the same layer output of the encoder are obtained through jump connection, and the two are spliced through channels to achieve cross-level feature fusion; finally, the model obtains the cumulative precipitation results for the next 1-6 weeks.

[0031] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0032] (1) In terms of data, the present invention considers the influence mechanism of multi-source data and integrates the data of polar vortex as an auxiliary to improve the accuracy of forecast correction;

[0033] (2) This invention adopts the channel and spatiotemporal attention mechanism to strengthen the weight of key meteorological factors, time and space, and ensure the effective use of historical meteorological data to the greatest extent;

[0034] (3) The present invention adopts a multi-scale feature extraction model. The convolution layer uses convolution kernels of different sizes for calculation, which expands the receptive field and improves the generalization ability of the model, thereby enhancing the robustness and adaptability of the model.

[0035] (4) The present invention selects the grid data with the strongest correlation when fusing polar vortex data. 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 impact of noise data and improves the calculation efficiency of the deep learning model. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a flow chart of a sub-seasonal precipitation forecast correction method based on MF-SPCM proposed in the present invention;

[0037] Figure 2 This is a model diagram of a sub-seasonal precipitation forecast correction method based on MF-SPCM proposed in the present invention;

[0038] Figure 3 This is a structural diagram of the TA model of the present invention;

[0039] Figure 4 Schematic diagram of the structure of spatial attention SA;

[0040] Figure 5 Schematic diagram of the SAConvLSTM model structure;

[0041] Figure 6 Schematic diagram of the MFCFN model structure;

[0042] Figure 7 Schematic diagram of the MCAF model structure;

[0043] Figure 8 This is a schematic diagram of the MDB module structure;

[0044] Figure 9 This is the result of the ablation experiment;

[0045] Figure 10 The figure is a comparison of experimental results. DETAILED DESCRIPTION

[0046] The technical solution of the present invention is further described in detail below with reference to the accompanying drawings:

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. The present invention is described in detail below based on the drawings and preferred embodiments. The purpose and effect of the present invention will become more clear. It should be understood that the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0048] The technical solution of the present invention will now be further described with reference to the accompanying drawings and embodiments.

[0049] Example 1:

[0050] like Figure 1 and Figure 2 As shown, this embodiment proposes a sub-seasonal precipitation forecast correction method based on MF-SPCM, which mainly includes the following steps:

[0051] Step 1: Obtain the S2S (Sub-seasonal to Seasonal) precipitation forecast for the target area for the next 1-6 weeks provided by the National Centers for Environmental Prediction (NCEP), ERA5 historical meteorological data, digital elevation model (DEM) data, and Arctic polar vortex data. The precipitation forecast data, historical meteorological data, DEM data, and polar vortex data are all gridded data. The spatial resolution of the precipitation forecast, historical meteorological, and polar vortex data is 0.25°, approximately 10 km, and the spatial resolution of the DEM data is 1 km. The precipitation forecast statistics are cumulative precipitation, i.e., the accumulated precipitation data from the forecast start date to the end date. The temporal resolution of the historical meteorological data is 1 hour. The polar vortex data are geopotential height data on the 10hPa, 50hPa, and 70hPa isopotential surfaces in the region 60°N-90°N and -180°E-180°E, with a temporal resolution of 1 hour. The specific operations include:

[0052] In this embodiment, the middle and lower reaches of the Yangtze River are located within the range of 24.5°-32°N and 110°-122°E. The historical meteorological data for the region from 2015 to 2024 were obtained from the ERA5 dataset released by the European Centre for Medium-Range Weather Forecasts (ECMWF). The temporal 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 obtained from the geospatial cloud platform with a spatial resolution of 30m. After resampling, the resolution is 1km. The slope and aspect data at the 1km resolution are then calculated. The polar vortex data is obtained from the ERA5 dataset, using the geopotential height data on the three equipotential surfaces of 10hPa, 50hPa, and 70hPa in the region of 60°N-90°N and -180°E-180°E.

[0053] Step 2: Preprocess the data. First, process the regional ERA5 historical meteorological data to obtain daily average meteorological data. The total precipitation element is calculated from the hourly precipitation statistics to the daily cumulative precipitation. When the total precipitation is used as the label, it is the cumulative precipitation from the forecast start day to the last day. The DEM data is calculated and processed to obtain elevation, slope, and aspect data. The polar vortex data is also preprocessed to obtain daily potential height values. Finally, all data are standardized. The specific operations include:

[0054] In this embodiment, the daily average values of the other 11 meteorological elements are first calculated. For the total precipitation, the daily cumulative precipitation is obtained. When the total precipitation is used as a label, the cumulative precipitation from the start day to the last day of the forecast is counted. The daily average value is also calculated for the polar vortex data. The formula used in the normalization method is:

[0055] (1)

[0056] Where, For input data, is the standardized data, is the mean of the input data, is the standard deviation of the input data.

[0057] Step 3: Input DEM (Digital Elevation Model) data into the Multi-Scale Feature Convolution Fusion Network (MFCFN) to obtain DEM features. Input the past 30 days of ERA5 meteorological data into the Temporal-Spatial Attention Convolutional LSTM (T-SAConvLSTM) model to obtain the spatiotemporal features of the early forecast period. Polar vortex data is mapped using teleconnection mapping to obtain a result consistent with the forecast data dimension. The T-SAConvLSTM model network is then used to further extract spatiotemporal features. Finally, a feature map of the multi-source data is obtained. The specific operations include:

[0058] In this embodiment, the MFCFN network is as follows: Figure 6As shown in the figure, the workflow is as follows: Initial terrain features are first obtained by inputting normalized elevation, slope, and aspect data into a residual block. This residual block consists of two 3×3 convolutional layers, two Bn layers, and two ReLU activation layers. Furthermore, the residual connection branch contains a 1×1 convolution and a Bn layer. To capture terrain features at multiple scales, the initial features are sequentially passed through three heterogeneous convolutional layers with kernel sizes of 7×7, 5×5, and 3×3, respectively. This design enables the model to extract terrain features at coarse, medium, and fine spatial scales. In the feature fusion stage, a two-stage strategy is employed for feature reconstruction. For example, at a layer with a 5×5 kernel size, the features of the previous layer are first upsampled to restore the spatial resolution. Simultaneously, the features of the current layer are reduced in channel dimension using 1×1 convolutions. Furthermore, pixel-wise addition is used to semantically align features at different levels, and channel concatenation is used to integrate the multi-scale feature maps. To enhance local feature representation and spatial resolution, this network integrates a series of standardized operation modules within each convolutional processing layer, including batch normalization, convolution operations, average pooling, and ReLU nonlinear activation functions. This modular design not only maintains a balance between network depth and computational efficiency, but also the multi-scale feature cross-fusion mechanism helps improve the extraction accuracy of fine-grained features such as terrain edges and textures, providing a robust terrain feature representation foundation for subsequent improvements in precipitation forecasting tasks. The final MFCFN model outputs a terrain feature map with dimensions of (3, 31, 49).

[0059] The TA model, SA model and SAConvLSTM network structures are respectively as follows Figure 3 、 Figure 4 and Figure 5As shown in Figure 2, the core algorithm of the TA model implements dynamic temporal weight modeling through multi-stage feature transformation. First, a channel-oriented global average pooling operation is performed on the input ERA5 meteorological data. To address the heterogeneous temporal influences of meteorological elements, this model performs independent channel GAP calculations on each of the 12 meteorological elements, reducing the original 12×30×31×49 four-dimensional tensor to a 12×30×1×1 matrix of initial time weights. This process extracts the initial significant features of each meteorological element at different time steps by performing mean operations along the spatial dimensions (H, W). However, traditional GAP, as a static calculation method, can only capture linear temporal correlations but has difficulty characterizing the nonlinear interactions of multiscale processes in the atmospheric system. To address the limitations of GAP, the present invention constructs a CMLP model after the GAP layer. This model adopts a hierarchical architecture of "grouping-dimensionality increase-nonlinear activation-dimensionality reduction." While ensuring computational efficiency, it achieves nonlinear mapping of the temporal feature space and mining of high-order interactions. Furthermore, it ensures that each meteorological element is not confounded by the temporal weights of other elements. Furthermore, this design ensures that the temporal weight of each meteorological factor is calculated independently, avoiding interference between different factors and thus balancing feature expressiveness with model generalization. The resulting 12×30×1×1 feature tensor after CMLP transformation is finally normalized along the temporal dimension using the Softmax function to generate dynamic temporal attention weights, thereby accurately quantifying the differences in the contribution of meteorological factors to precipitation forecasts at different historical moments. The calculation formula is as follows:

[0060] (2)

[0061] in, stands for element-wise multiplication, and Represents input data and output data respectively.

[0062] The SA-ConvLSTM model optimizes the spatiotemporal feature extraction capability of the traditional ConvLSTM by embedding the spatial attention mechanism. Its core structure is as follows: Figure 5 As shown. This model integrates the spatial attention mechanism at the input of ConvLSTM, which inputs the current meteorological elements into the data and the hidden state at the previous moment After channel splicing, the input is input into the spatial attention model for calculation. The spatial attention model SA contains maximum pooling, average pooling and Softmax functions. The structure is as follows Figure 4As shown in the figure, these calculations generate a feature map with a spatial weight distribution. This design enables the model to dynamically calculate the importance weights of different meteorological elements at each grid point, accurately weighting features in key spatial regions, and effectively enhancing the ConvLSTM's ability to capture spatial features in complex meteorological fields. Finally, the model performs dimensionality reduction on the extracted spatiotemporal features of the 12 meteorological elements, ultimately outputting a 12×31×49 ERA5 historical meteorological element feature map.

[0063] The ConvLSTM calculation process is shown in Equations (3) to (7).

[0064] (3)

[0065] (4)

[0066] (5)

[0067] (6)

[0068] (7)

[0069] in, Represents the space-time meteorological field input at time t. Input gate , Forget Gate and output gate The gating coefficients in the range of 0 to 1 are dynamically generated through convolution operation and Sigmoid activation function. The input gate is used to update the cell state, the forget gate is used to select the meteorological information input at time t to be forgotten, and the output gate is used to control the spatiotemporal characteristics of the current output. These gating mechanisms are obtained by calculating the corresponding gating coefficients based on the input at the current time t and the cell state at the previous time t-1, 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. Cell state The forget gate and the input gate are jointly regulated to realize the selective memory of historical information, as shown in Equation 8, where is element-wise multiplication, Long-term memory of the previous moment; output gate Regulate the information output of the cell state to generate the current hidden state , where tanh represents the tanh activation function.

[0070] The technical process of the teleconnection mapping method can be decomposed into three core steps: first, a null value tensor matrix is initialized as the correlation field container based on the spatial dimension of the forecast data. The dimension of the null value tensor matrix is (30, 3, 31, 49); second, the temporal correlation matrix between the ERA5 precipitation field and the polar vortex field is calculated grid by grid point using the Pearson correlation coefficient algorithm, and the spatial coordinate point with the maximum correlation coefficient in the polar vortex field is identified; finally, the potential height value of the optimal correlation point is mapped to the corresponding ERA5 geographic spatial 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).

[0071] (8)

[0072] (9)

[0073] (10)

[0074] (11)

[0075] in, is the empty tensor matrix created, represents an empty tensor matrix, For the ERA5 precipitation field, It is the polar vortex field. is the result of Pearson correlation coefficient calculation, is the covariance, is the standard deviation; is the polar vortex field grid point position with maximum correlation, is the value of the polar vortex field at that particular location; It is a data carrier that carries key information of the polar vortex field.

[0076] After teleconnection mapping, the polar vortex data is transformed from its original dimensions of (30, 3, 121, 1440) to gridded data with a stronger correlation with precipitation, with dimensions of (30, 3, 31, 49). The spatiotemporal features of this tensor are then further extracted using the aforementioned T-SAConvLSTM (Temporal-Spatial Attention Convolutional LSTM) network model, resulting in a spatiotemporal feature map with dimensions of (3, 31, 49).

[0077] Step 4: Input the multi-source data feature map and the precipitation forecast data to be corrected into the Multi-source Cross-Attention Fusion (MCAF) module to obtain the weighted multi-source data feature map;

[0078] 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 is modeled separately through a 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:

[0079] (12)

[0080] (13)

[0081] (14)

[0082] in, is the forecast query vector, is the prediction key vector, is the forecast value vector; is the key vector dimension, which is the 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 sum to 1; Represents the predicted attention output matrix, the final feature representation obtained by 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 splicing operation; is the fusion attention matrix.

[0083] Step 5: Input the weighted multi-source data feature map into the MDBU-Net correction module that integrates the Maxpool and Dense Block (MDB) to obtain the revised sub-seasonal cumulative precipitation forecast for the next 1-6 weeks.

[0084] The MDB module structure is as follows Figure 8 As shown in the figure, this module uses a max-pooling operation to compress the spatial dimension of the input feature map (downsampling it to 1 / 4 of its original size), preserving the salient features of meteorological elements while suppressing noise interference. It then constructs a convolutional block consisting of three densely connected layers. The input to each convolutional layer consists of the output feature maps of all preceding layers concatenated along the channel dimension. Specifically, the number of input channels in the nth layer is the sum of the initial input channels and the output channels of the n-1th layer. This dense feed-forward feature significantly enhances gradient flow efficiency and feature representation by establishing a cross-layer feature reuse mechanism. During the encoding phase of MDBU-Net, this cross-layer gradient propagation mechanism is used to deeply mine the spatiotemporal correlations of the fused feature maps from multi-source data. Finally, the coarse-grained features output by the max-pooling path are concatenated along the channel dimension with the fine-grained features extracted by the dense convolutional path, forming a composite feature representation that is both spatially robust and detail-resolving. This dual-stream fusion architecture, through the synergistic mechanism of feature selection and enhancement, effectively mitigates the feature resolution degradation caused by traditional downsampling while maintaining computational efficiency.

[0085] During the decoding phase, when upsampling is achieved through deconvolution, deconvolution is first performed on the low-resolution feature map to restore the spatial resolution of the feature map to the size of the corresponding encoder layer. Skip connections are then used to obtain multi-scale features from the encoder's output layer. These features are then concatenated across channels to achieve cross-layer feature fusion. Finally, the model generates cumulative precipitation results for the next 1-6 weeks.

[0086] The relevant data from May to August from 2015 to 2023 are used as the training set. The training labels are the cumulative precipitation values for the next 1 to 6 weeks. The Adam optimizer is used to train the model with a learning rate of 0.001 and a loss function using the mean absolute error (MAE). The formula is:

[0087] (15)

[0088] in, is the number of samples, It is The true value of the sample, It is The forecast correction value of the sample.

[0089] The trained model is used to predict the test set data to determine the accuracy of the model.

[0090] This example sets up an ablation experiment to verify the effectiveness of each module, and through comparative experiments, it proves the accuracy advantage of the proposed model compared with existing advanced methods. During the experiment, the root mean square error (RMSE) is used as the evaluation indicator. The calculation formula is as follows:

[0091] (16)

[0092] in 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 of the revised model, The first samples, is the first grid point data in the correction area. grid points.

[0093] First, we conduct an ablation experiment. The ablation experiment results from May to August 2024 are shown in Table 1. The quantitative results are shown in Table 1, and the visualization results of the RMSE distribution are shown in Table 1. Figure 9 As shown in Table 1, it can be clearly seen 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 the multi-source data modules are gradually integrated, indicating that these modules play a positive role. Figure 9 As shown in the figure, MF-SPCM exhibits the best error performance. With the gradual introduction of different data, the RMSE indicator of the model shows a significant downward trend. This phenomenon not only verifies the role of incremental data in improving model performance, but also highlights the effectiveness of data quality and model design.

[0094] Table 1

[0095] method MFCFN T-SAConvLSTM Teleconnection mapping MCAF MDB RMSE NCEP forecast results 123.9864 DEM ✔ 86.3633 DEM-ERA5 ✔ ✔ 85.0876 Multi-source data ✔ ✔ ✔ 83.2892 MCAF ✔ ✔ ✔ ✔ 82.5379 MF-SPCM ✔ ✔ ✔ ✔ ✔ 81.6757

[0096] In the comparative experiment, the NCEP forecast results were compared with the revised results of multiple advanced models A-ConvLSTM, PDFN, CSG-UNET and EFC-Net. The comparative experiment results from May to August 2024 are shown in Table 2. The quantitative results are shown in Table 2, and the visual results are shown in Table 2. Figure 10 As shown in the figures and tables, it can be seen that compared with these advanced models, the MF-SPCM proposed in this invention has obvious advantages and shows the lowest error value of the correction results.

[0097] Table 2

[0098] index NCEP A-ConvLSTM PDFN CSG-UNET EFC-Net MF-SPCM RMSE 123.9864 93.5689 84.6914 94.3823 87.0681 81.6757

[0099] The present invention proposes a deep learning method for sub-seasonal precipitation forecast correction based on multi-source data fusion. In terms of data, it includes terrain data and meteorological element data of the region, and innovatively integrates polar vortex data. In terms of method, different neural network modules are used to process multi-source data, and at the same time, feature information beneficial to precipitation forecast contained in these data is extracted, thereby improving the accuracy of precipitation forecast. In the backbone network of the end-to-end model, the improved MDBU-Net is used to perform deep and complex calculations. The model adopts the residual construction idea to reduce the information loss of the network during the operation process.

[0100] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the above-described method embodiments. Any reference to memory, storage, database, or other media used in the embodiments provided herein may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage. Volatile memory may 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).

[0101] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0102] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A method for correcting subseasonal precipitation forecasts based on MF-SPCM, characterized by: The specific steps include: Step 1: Obtain the sub-seasonal to seasonal (S2S) precipitation forecast results for the target area for the next 1-6 weeks provided by the National Centers for Environmental Prediction (NCEP), ERA5 historical meteorological data, digital elevation model (DEM) data, and Arctic polar vortex related data; Step 2: Data preprocessing: First, the regional ERA5 historical meteorological data is processed to obtain daily mean meteorological data. The total precipitation element is calculated from hourly precipitation statistics to daily cumulative precipitation. When total precipitation is used as a label, it is the cumulative precipitation from the forecast start day to the last day. Digital elevation model (DEM) data is calculated and processed to obtain elevation, slope, and aspect data. Arctic polar vortex data is also preprocessed to obtain daily potential height values. Finally, all data are standardized. Step 3: Input the DEM related data into the multi-scale feature convolution fusion network MFCFN to obtain the DEM features; The past 30 days of ERA5 historical meteorological data were input into the spatiotemporal attention convolutional long short-term memory (T-SAConvLSTM) model to obtain the spatiotemporal features of the forecast period. Polar vortex data were mapped using teleconnection mapping to obtain a mapping result consistent with the forecast data dimension. The T-SAConvLSTM model network was then used to further extract spatiotemporal features, ultimately resulting in a 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, which integrates the maximum pooling and dense connection block MDB, to obtain the revised results of the sub-seasonal cumulative precipitation forecast for the next 1-6 weeks.

2. The method for correcting subseasonal precipitation forecast based on MF-SPCM according to claim 1, characterized in that: In step 1, precipitation forecast data, historical meteorological data, digital elevation model (DEM) data, and Arctic polar vortex data are all grid data; the spatial resolution of precipitation forecast data, historical meteorological data, and Arctic polar vortex data is 0.25°, 10 km, and the spatial resolution of DEM data is 1 km; precipitation forecast statistics are cumulative precipitation, that is, the cumulative precipitation data from the forecast start date to the end date, and the time resolution of historical meteorological data is 1 hour; Arctic polar vortex data are geopotential height data on 10hPa, 50hPa, and 70hPa isopotential surfaces in the 60°N-90°N, -180°E-180°E area, with a time resolution of 1 hour.

3. The method for correcting subseasonal precipitation forecast 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, achieving collaborative representation learning of multi-scale features through multi-branch convolution paths and cross-layer connection mechanisms. Specifically, the input three-channel DEM raw data including elevation, slope, and aspect are processed by a deep feature extraction unit containing residual blocks to generate initial DEM features. The residual block consists of two 3×3 convolutional layers, two Bn layers, and two ReLU activation function layers. The residual connection branch also consists of a 1×1 convolutional layer and a Bn layer. To capture the multi-scale geometric features of the terrain data, three sets of heterogeneous convolutional kernels are designed in the network, with sizes of 7×7, 5×5, and 3×3, respectively. This allows the model to extract terrain feature information at coarse, medium, and fine spatial scales, respectively. In the feature fusion stage, a two-stage strategy is adopted for feature reconstruction. Taking the layer with a convolution kernel size of 5×5 as an example, the spatial resolution of the previous layer features is restored through upsampling operations, and the channel dimension of the current layer features is reduced using 1×1 convolution. The semantic alignment of features at different levels is achieved through pixel-level addition, and the multi-scale feature maps are integrated through channel splicing operations.

4. The method for correcting subseasonal precipitation forecast based on MF-SPCM according to claim 1, characterized in that: In step 3, the T-SAConvLSTM model network is implemented by a serial architecture. The ERA5 historical meteorological data includes 2m temperature, 10m wind field, and humidity meteorological elements. The time period is from 30 days before the forecast start date to 1 day before the forecast start date, a total of 30 days of historical meteorological data. 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 is first extracted from the temporal features by GAP, and then nonlinearly transformed by CMLP. Then, a time attention weight matrix with physical significance is generated by 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 the SAConvLSTM network model. Finally, the two are combined to construct a spatiotemporal feature extractor with physical interpretability. The output data of the previous module is input into the SAConvLSTM model, and finally the historical meteorological element feature map is obtained. The calculation process is shown in formula (1): (1) in, stands for element-wise multiplication, and Represents input data and output data respectively.

5. The method for correcting subseasonal precipitation forecast 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 spatiotemporal correlations of the forecast data and multi-source data features are modeled separately through a parallel self-attention mechanism. The forecast branch uses the self-attention layer to capture the potential cumulative relationship between each forecast period, while the auxiliary feature branch uses self-attention to enhance the spatial correlation representation of terrain and circulation elements. Subsequently, the two sets 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 as key-value pairs. A cross-modal dynamic feature mapping relationship is established through a learnable attention weight matrix, realizing 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 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: (2) (3) (4) in, is the forecast query vector, is the prediction key vector, is the forecast value vector; is the key vector dimension, which is the 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 sum to 1; Represents the predicted attention output matrix, the final feature representation obtained by 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 splicing operation; is the fusion attention matrix.

6. The method for correcting subseasonal precipitation forecast based on MF-SPCM according to claim 1, characterized in that: In step 3, the teleconnection mapping method consists of three steps: (1) Initialize a null value tensor matrix as the correlation field container based on the spatial dimension of the forecast data; (2) The temporal correlation matrix between the ERA5 precipitation field and the polar vortex field during the same period was calculated grid by grid using the Pearson correlation coefficient algorithm, and the spatial coordinate point with the maximum correlation coefficient in the polar vortex field was identified; (3) Map the geopotential height value of the optimal correlation point to the corresponding ERA5 geographic spatial position in the empty matrix. The resulting mapping matrix is the mapping result of the teleconnection. The calculation process is shown in Equations (5) to (8): (5) (6) (7) (8) in, is the empty tensor matrix created, represents an empty tensor matrix, For the ERA5 precipitation field, is the polar vortex field, is the result of Pearson correlation coefficient calculation, is the covariance, is the standard deviation; is the polar vortex field grid point position with maximum correlation, is the value of the polar vortex field at that particular location; It is a data carrier that carries key information of the polar vortex field.

7. The method for correcting subseasonal precipitation forecast based on MF-SPCM according to claim 1, characterized in that: In step 5, the correction module MDBU-Ne uses a maximum pooling operation to perform spatial dimensional compression on the input feature map, suppressing noise interference while retaining the significant features of meteorological elements. It then constructs a convolution block with a three-layer dense connection structure, in which the input of each convolution operation is composed of the output feature maps of all previous layers spliced 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-1th layer. The encoding stage deeply mines the spatiotemporal 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 maximum 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 capabilities. In the decoding stage, when upsampling is achieved through deconvolution, deconvolution is first performed 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, the multi-scale features of the same-layer output of the encoder are obtained through jump connections, and the two are combined through channel splicing to achieve cross-level feature fusion; finally, the model obtains the cumulative precipitation results for the next 1-6 weeks.

Images