Typhoon rainfall intelligent prediction method based on multi-scale space-time coupling

Abstract

The invention discloses a typhoon rainfall intelligent prediction method based on multi-scale space-time coupling, comprising the following steps: 1) collecting typhoon rainfall data including satellite rainfall data, numerical weather forecast data and typhoon path information; 2) data preprocessing; comprising data standardization and space-time cutting; 3) data coding; 4) extracting spatial features; spatial feature extraction is carried out by adopting a double-branch spatial coding module and is divided into a local convolution branch LC and a global Fourier transform branch GF; 5) multi-scale coupling; and 6) decoding by using a decoder, and outputting a typhoon rainfall field prediction result. The hidden representation output by the space coding module is used as an input to obtain a hidden feature of a next time step. According to the typhoon rainfall prediction model constructed by the method, the modeling capability and the prediction precision of the typhoon rainfall process are remarkably improved by constructing the adaptive spatial codes and combining the multi-scale spatial-temporal feature extraction structure, the flexible decoder and the regional loss function.

Description

Technical Field

[0001] This invention relates to meteorological disaster prediction technology, and more particularly to an intelligent prediction method for typhoon rainfall based on multi-scale spatiotemporal coupling. Background Technology

[0002] Typhoons and torrential rains are the main forms of typhoon disasters, with far-reaching impacts and complex, variable natures. They frequently lead to secondary geological disasters such as floods, landslides, and mudslides, causing significant impacts on human society. Therefore, accurately predicting the spatial distribution and evolution of rainfall during typhoons is crucial for disaster prevention, mitigation, and public early warning response. Currently widely used numerical weather prediction (NWP) models, such as WRF and ECMWF, perform well in simulating large-scale weather systems, but their high sensitivity to initial conditions and high computational resource consumption make it difficult to achieve high-frequency updates and high-resolution output. Furthermore, single deep learning models have limited predictive capabilities when dealing with rapidly changing nonlinear systems like typhoons. To address these issues, there is an urgent need to construct a typhoon rainfall prediction model that considers local details, global structure, and long- and short-term evolution processes, possessing strong generalization capabilities, high accuracy, and flexible output capabilities to meet the practical needs of current intelligent meteorological services and disaster prevention and mitigation operations. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a smart typhoon rainfall prediction method based on multi-scale spatiotemporal coupling, which addresses the shortcomings of the existing technology.

[0004] The technical solution adopted by this invention to solve its technical problem is: a smart typhoon rainfall prediction method based on multi-scale spatiotemporal coupling, comprising the following steps: 1) Collect typhoon rainfall data, including satellite rainfall data, numerical weather prediction data, and typhoon track information; Among them, satellite precipitation data is a spatially continuous image of the typhoon's precipitation field; numerical weather prediction data includes the typhoon's wind speed and geopotential height; typhoon track information includes the latitude and longitude of the typhoon center, as well as the maximum wind speed and minimum air pressure at the center; 2) Data preprocessing; including data standardization and spatiotemporal pruning; 3) Data encoding; After aligning all the preprocessed data along the time dimension, they are uniformly processed into a five-dimensional tensor format: X=(B,T,C,H,W), where B represents the data batch size, T is the time step, C represents the number of channels, H is the spatial height of the data, and W is the spatial width of the data. 4) Spatial feature extraction; Spatial feature extraction is performed using a dual-branch spatial coding module, which consists of a local convolution branch (LC) and a global Fourier transform branch (GF). Among them, the local convolution branch is based on a shallow convolutional neural network, which extracts feature maps of high-frequency detail features from the input image through multiple convolution and downsampling operations; Global frequency domain branch; used to capture large-scale spatial dependencies and global correlations in the rainfall field based on the input image; The two branches interact through multiple iterations to achieve the fusion of dual-branch features; 5) Multi-scale coupling; using the fused dual-branch features output by the spatial coding module as input, the hidden features for the next time step are obtained: 6) Use a decoder to decode and output the typhoon rainfall field prediction results.

[0005] According to the above scheme, in step 2), the spatiotemporal clipping adopts a dynamic clipping strategy with the typhoon center as the anchor point. With the typhoon center at each time step as the anchor point, a region of fixed spatial size is clipped from the original image and weather forecast variable field corresponding to the satellite rainfall data. The region covers the typhoon core and its surrounding area, ensuring the spatial and physical alignment of rainfall data and meteorological variables.

[0006] According to the above scheme, in step 3), the data encoding is as follows; 3.1) Satellite rainfall data processing: The cropped multiple frames of images are stacked in the time dimension to form an image tensor with a time series structure; 3.2) Numerical meteorological variable processing: After spatial clipping and layer-by-layer normalization, the numerical meteorological variables are consistent with the precipitation image in spatial dimension, and are stacked into tensors after being aligned in the time dimension; 3.3) Typhoon path information processing: The discrete point data of typhoon path information is visualized to form a channel with the same structure as satellite precipitation data images; 3.4) Multimodal tensor fusion representation: After all preprocessed data are aligned in the time dimension, they are uniformly processed into a five-dimensional tensor format: X=(B, T, C, H, W), where B represents the data batch size, T is the time step, C represents the number of channels, H is the spatial height of the data, and W is the spatial width of the data.

[0007] According to the above scheme, in step 4), the LC branch is... It consists of stacked ConvNormReLU units: its core structure is as follows: ; ; Where GNorm represents group normalization, and LeakyReLU is the Leaky ReLU activation function. This represents the output of the i-th LC block; This represents a two-dimensional convolutional layer.

[0008] According to the above scheme, step 4) of the global Fourier transform branch processing includes: Step 4.1) Divide the input image into non-overlapping image patches and linearly map them into embedding vectors; Step 4.2) For the obtained image patch embedding vector Spatial domain features are converted into frequency domain features through two-dimensional fast Fourier transform (2D FFT);

[0009] Where l represents frequency and x represents spatial position. This represents the frequency domain representation updated by the Fourier transform; Step 4.3) Use a multilayer feedforward neural network (MLP) to perform a linear transformation on the frequency domain features, mapping them to a linear space; Step 4.4) The processed frequency domain features are transformed back into the spatial domain through inverse Fourier transform (IFFT) to achieve the combination of frequency domain information and spatial details; ; in, This represents the result of a linear transformation.

[0010] According to the above scheme, in step 5), the decoder includes a spatial decoder and a temporal decoder; Among them, the spatial decoder is composed of Composed of several decoding blocks, the spatial resolution of the encoded input features is gradually restored through a deconvolution (ConvTranspose2D) layer, mapping the low-dimensional spatiotemporal features back to the original rainfall field size, ensuring high-resolution output of the predicted image; ; .

[0011] According to the above scheme, in step 5), the time decoder uses ConvNormReLU units to extend the time channels. Specifically, the features obtained from the first stage decoding... The time dimension and channel dimension are concatenated to form a tensor of size T×C, which is then mapped to K×C, where K is the target prediction length. Subsequently, the resulting feature map is subjected to dimensionality transformation to obtain the target prediction dimension. ; The calculation process is as follows: .

[0012] According to the above scheme, the loss function is calculated separately for the typhoon's core region. , peripheral area and the entire region The error in precipitation forecasting.

[0013] set up These represent the predicted and actual precipitation fields, respectively. A binary mask for the cyclone's core region. It serves as a mask for the outer area.

[0014] The regional MSE loss function is defined as follows: ; Total loss function: ; in, This represents the actual precipitation of the b-th sample at time t. A binary mask for the kernel region. For the outer area; , and These are the weighting coefficients.

[0015] The beneficial effects of this invention are: 1. This invention addresses the complex dynamics of rainfall time series data, which exhibit both short-term volatility and long-term trends, by designing a multi-scale temporal convolution and Fourier coupling mechanism. The multi-scale features of the time series are implemented through a multi-layered convolutional network (TCN). Its core advantage is its ability to process long-sequence data in parallel, avoiding the gradient vanishing problem inherent in recurrent neural networks (RNNs), and supporting flexible adjustment of the receptive field to capture changes at different time scales.

[0016] 2. In the frequency domain analysis module of the time series, this invention utilizes one-dimensional fast Fourier transform (1D FFT) to periodically decompose the time dimension features, extracting potential meteorological cycle patterns and seasonal variation trends. The frequency domain time features are then fused with convolutional time features after nonlinear mapping, forming a unified spatiotemporal coupled representation. This spatiotemporal frequency domain fusion mechanism not only effectively improves the understanding of complex rainfall evolution patterns but also significantly alleviates the problems of error accumulation and prediction divergence in long-term multi-step predictions, ensuring the continuity and structural consistency of the prediction sequence.

[0017] 3. This invention constructs a typhoon rainfall prediction model that takes into account local details, global structure, and long- and short-term evolution processes. By constructing an adaptive spatial encoding, combined with a multi-scale spatiotemporal feature extraction structure, a flexible decoder, and a regional loss function, it significantly improves the modeling ability and prediction accuracy of typhoon rainfall processes. The model of this invention has strong generalization ability, predictive ability, and flexible output characteristics to meet the actual needs of current intelligent meteorological services and disaster prevention and mitigation operations. Attached Figure Description

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of a method according to an embodiment of the present invention; Figure 2 This is a model structure diagram of an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] like Figure 1 As shown, a smart typhoon rainfall prediction method based on multi-scale spatiotemporal coupling includes the following steps: 1) Collect typhoon rainfall data, including satellite rainfall data, numerical weather prediction data, and typhoon track information; Among them, satellite precipitation data is a spatially continuous image of the typhoon's precipitation field; numerical weather prediction data includes the typhoon's wind speed and geopotential height; and typhoon track information is the latitude and longitude of the typhoon's core. In this embodiment, the collected data includes satellite precipitation data (CMORPH), ERA5 reanalysis data, and typhoon optimal path data (Northwest Pacific region) from 1998 to 2021. The data used included hourly CMORPH data with a spatial resolution of 0.25°×0.25°, covering a global region from 60°S to 60°N; ERA5 reanalysis data selected meteorological variables from multiple pressure layers of ERA5 provided by ECMWF, including horizontal wind components (u, v) and geopotential height (z), with a spatial resolution of 0.25° and a temporal resolution of 1 hour; and typhoon optimal track data including the typhoon's latitude and longitude, maximum wind speed at its center, and minimum air pressure every 6 hours. Dataset partitioning: Typhoons in the Northwest Pacific region from 1998 to 2021 were randomly divided into three subsets: the training set (70% of the total sample size) was used for model parameter learning; the validation set (10%) was used for hyperparameter tuning and overfit monitoring; and the test set (20%) was used to evaluate the model's generalization performance on unseen samples.

[0021] 2) Data preprocessing; including data standardization and spatiotemporal pruning; 2.1) A dynamic cropping strategy with the typhoon center as the anchor point is adopted: the typhoon center at each time step is used as the anchor point to crop a region of fixed spatial size from the original image or weather forecast variable field corresponding to the satellite precipitation data. The region covers the typhoon core and its surrounding area to ensure the spatial and physical alignment of precipitation data and meteorological variables. 2.2) Standardization: Local normalization is performed on each data channel; 3) Data encoding; 3.1) Satellite rainfall data processing: The cropped multiple frames of images are stacked in the time dimension to form an image tensor with a time series structure; 3.2) Numerical meteorological variables: After spatial clipping and layer-by-layer normalization, the numerical meteorological variables are consistent with the precipitation image in spatial dimension and stacked into tensors after being aligned in the time dimension; 3.3) Typhoon path information: Visualize the discrete point data (latitude and longitude, wind speed, etc.) of typhoon path information to form a channel with the same structure as satellite precipitation data images; Specifically, the location and intensity information of the typhoon are assigned to four blank grid cells. Each grid cell extends outward from the center point, covering an area of ​​a×a (where a=7). Pixels outside this area are uniformly set to the default value of 0. 3.4) Multimodal tensor fusion representation: After all preprocessed data are aligned in the time dimension, they are uniformly processed into a five-dimensional tensor format: (B, T, C, H, W), where B represents the data batch size, T is the time step, C represents the number of channels, H is the spatial height of the data, and W is the spatial width of the data. 4) Spatial feature extraction; Spatial feature extraction employs a dual-branch spatial coding module, which consists of a local convolution branch and a global Fourier transform branch. Before the local and global branches are executed in parallel, the multi-source inputs are first preprocessed through a Stem module, which is composed of consecutive 1×1 convolutions for preliminary feature extraction. Among them, the local convolution branch is based on a shallow convolutional neural network. Through multi-layer convolution and downsampling operations, it extracts feature maps of high-frequency detail features from the input image (such as edges, textures, and small-scale high-gradient spaces including the eyewall of a typhoon, spiral rainbands, and rainfall fronts).

[0022] By introducing batch normalization and the ReLU activation function, we can enhance the nonlinear fitting capability and avoid gradient problems.

[0023] In this embodiment, the LC branch is... It consists of stacked ConvNormReLU units: its core structure is as follows: ; ; Where GNorm represents group normalization, and LeakyReLU is the Leaky ReLU activation function. This represents the output of the i-th LC block; Represents a two-dimensional convolutional layer; The core function of the LC branch is to maintain the sharpness and structural clarity of local features and enhance sensitivity to complex texture changes.

[0024] Global Fourier transform branch; used to capture large-scale spatial correlations in rainfall fields through frequency domain transformation; The global Fourier transform branch and the local convolution branch run in parallel, and at each stage, the feature representation maintains the same spatial size as the output of the local branch.

[0025] include: Step 1: Divide the input image into non-overlapping image patches and linearly map them into embedding vectors; set up Given the input image at time t, it is first divided into... Each image patch consists of non-overlapping p×p image blocks. Each image block is projected onto an embedding vector through a linear layer. This allows us to obtain a tokenized representation of the image.

[0026] The (;...;) indicates a row-by-row stacking operation.

[0027] Step 2: Convert the spatial domain features into frequency domain features using a two-dimensional fast Fourier transform (2D FFT); For the obtained image patch embedding vector Perform a Fourier transform to obtain the frequency domain representation:

[0028] Where l represents frequency and x represents spatial position. This represents the frequency domain representation updated by the Fourier transform; Step 3: Use a multilayer feedforward neural network (MLP) to perform a nonlinear transformation on the frequency domain features, mapping them to a linear space; ; Step 4: Transform the processed frequency domain features back into the spatial domain using the inverse Fourier transform (IFFT) to achieve the combination of frequency domain information and spatial details; ; in, This represents the result of a linear transformation; Global-local interaction: Both branches use two-dimensional convolutional layers for upsampling and downsampling operations, and achieve the fusion of features between the two branches through multiple iterative interactions. 5) Multi-scale coupling; using the hidden representation output by the spatial coding module. (Fused features) are used as input to obtain the hidden features at the next time step. : ; ; Where Ne is the number of temporal blocks. The hidden representation output by the spatial coding module; 6) Use a decoder to decode and output the typhoon rainfall field prediction results. The decoding stage includes a spatial decoder and a temporal decoder; Spatial decoder, by Composed of several decoding blocks, the spatial resolution of the encoded input features is gradually restored through a deconvolution (ConvTranspose2D) layer, mapping the low-dimensional spatiotemporal features back to the original rainfall field size, ensuring high-resolution output of the predicted image; ; ; The time decoder uses ConvNormReLU units to extend the time channels. Specifically, it uses the features obtained from the first stage decoding... The time dimension and channel dimension are concatenated to form a tensor of size T×C, which is then mapped to K×C, where K is the target prediction length. Subsequently, the resulting feature map undergoes a dimensionality transformation to obtain the target prediction dimension. The calculation process is as follows: ; The structural diagram of the model of this invention is as follows: Figure 2 ; In the model, to more accurately assess the spatial distribution characteristics of typhoon precipitation forecasts, this invention designs a Regional Mean Squared Error (MSE) loss function. This loss function calculates the regional mean squared error (MSE) loss function for the typhoon core area. , peripheral area and the entire region The error in precipitation forecasting.

[0029] set up These represent the predicted and actual precipitation fields, respectively. A binary mask for the cyclone's core region. It serves as a mask for the outer area.

[0030] The regional MSE loss function is defined as follows: ; Total loss function: ; in, This represents the actual precipitation of the b-th sample at time t. A binary mask for the kernel region. For the outer area; , and These are the weighting coefficients.

[0031] By distinguishing errors in different regions, this loss function can better capture spatially uneven prediction biases, thereby enhancing the model's sensitivity to the spatial structure of precipitation fields. In this embodiment, the core region is defined as a circular area with a radius of 5 grid cells centered on the cyclone center. The set loss radius is slightly larger than the typical cyclone core scale and the 7×7 encoding area, ensuring that this range can completely cover the typhoon core region, i.e., the area where strong convection and heavy precipitation mainly occur. This design allows the encoded dynamic information to highly overlap spatially with the more constrained loss region, thereby improving the spatial consistency between feature representation and evaluation target.

[0032] To evaluate the model's ability to detect typhoon precipitation events, the F1 score was used as the primary classification indicator. The indicator was evaluated at multiple precipitation intensity thresholds: 0.5, 2, 5, 10, 20, and 30 mm / h.

[0033] The F1 score balances precision and recall: precision measures how many of the samples predicted as precipitation events are actually real events, while recall measures how many of the actual precipitation events are successfully identified by the model. The F1 score is the harmonic mean of the two, providing a comprehensive reflection of the model's ability to detect precipitation events, and is particularly suitable for situations with imbalanced sample classes.

[0034]

[0035] Hit (TP), False Alarm (FP), False Negative (FN), and True Rejection (TN) correspond to true positives, false positives, false negatives, and true negatives, respectively, forming the basis for calculating various test indicators. Table 1 compares the average F1 score of the model used in this invention with the ablation-corrected model configuration and the 2D CNN model on the test set across multiple thresholds.

[0036] Table 1. Comparison of average F1 scores for multiple thresholds based on the test set.

[0037] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A smart typhoon rainfall prediction method based on multi-scale spatiotemporal coupling, characterized in that, Includes the following steps: 1) Collect typhoon rainfall data, including satellite rainfall data, numerical weather prediction data, and typhoon track information; Among them, satellite precipitation data is a spatially continuous image of the typhoon's precipitation field; numerical weather prediction data includes the typhoon's wind speed and geopotential height; typhoon track information includes the latitude and longitude of the typhoon center, as well as the maximum wind speed and minimum air pressure at the center; 2) Data preprocessing; including data standardization and spatiotemporal pruning; 3) Data encoding; After aligning all the preprocessed data along the time dimension, they are uniformly processed into a five-dimensional tensor format: X=(B, T, C,H, W), where B represents the data batch size, T is the time step, C represents the number of channels, H is the spatial height of the data, and W is the spatial width of the data. 4) Spatial feature extraction; Spatial feature extraction is performed using a dual-branch spatial coding module, which consists of a local convolution branch (LC) and a global Fourier transform branch (GF). Among them, the local convolution branch is based on a shallow convolutional neural network, which extracts feature maps of high-frequency detail features from the input image through multiple convolution and downsampling operations; Global frequency domain branch; used to capture large-scale spatial dependencies and global correlations in the rainfall field based on the input image; The two branches interact through multiple iterations to achieve the fusion of dual-branch features; 5) Multi-scale coupling; using the fused dual-branch features output by the spatial coding module as input, the hidden features for the next time step are obtained: 6) Use a decoder to decode and output the typhoon rainfall field prediction results.

2. The intelligent typhoon rainfall prediction method based on multi-scale spatiotemporal coupling according to claim 1, characterized in that, In step 2), the spatiotemporal clipping adopts a dynamic clipping strategy with the typhoon center as the anchor point. With the typhoon center at each time step as the anchor point, a region of fixed spatial size is clipped from the original image and weather forecast variable field corresponding to the satellite rainfall data. The region covers the typhoon core and its surrounding area, ensuring the spatial and physical alignment of rainfall data and meteorological variables.

3. The intelligent typhoon rainfall prediction method based on multi-scale spatiotemporal coupling according to claim 1, characterized in that, In step 3), the data encoding is specifically as follows; 3.1) Satellite rainfall data processing: The cropped multiple frames of images are stacked in the time dimension to form an image tensor with a time series structure; 3.2) Numerical meteorological variable processing: After spatial clipping and layer-by-layer normalization, the numerical meteorological variables are consistent with the precipitation image in spatial dimension, and are stacked into tensors after being aligned in the time dimension; 3.3) Typhoon path information processing: The discrete point data of typhoon path information is visualized to form a channel with the same structure as satellite precipitation data images; 3.4) Multimodal tensor fusion representation: After all preprocessed data are aligned in the time dimension, they are uniformly processed into a five-dimensional tensor format: X=(B, T, C, H, W), where B represents the data batch size, T is the time step, C represents the number of channels, H is the spatial height of the data, and W is the spatial width of the data.

4. The intelligent typhoon rainfall prediction method based on multi-scale spatiotemporal coupling according to claim 1, characterized in that, In step 4), the local convolution branch is... It consists of stacked ConvNormReLU units: its core structure is as follows: ； ； Where GNorm represents group normalization, and LeakyReLU is the Leaky ReLU activation function. This represents the output of the i-th LC block; This represents a two-dimensional convolutional layer; D represents the number of feature channels.

5. The intelligent typhoon rainfall prediction method based on multi-scale spatiotemporal coupling according to claim 1, characterized in that, In step 4), the global Fourier transform branching process includes: Step 4.1) Divide the input image into non-overlapping image patches and linearly map them into embedding vectors; Step 4.2) For the obtained image patch embedding vector The spatial domain features are converted into frequency domain features through two-dimensional fast Fourier transform; Step 4.3) Use a multilayer feedforward neural network (MLP) to perform a linear transformation on the frequency domain features, mapping them to a linear space; Step 4.4) The processed frequency domain features are transformed back into the spatial domain by inverse Fourier transform (IFFT), thus combining frequency domain information with spatial details.

6. The intelligent typhoon rainfall prediction method based on multi-scale spatiotemporal coupling according to claim 1, characterized in that, In step 5), the decoder includes a spatial decoder and a temporal decoder; Among them, the spatial decoder is composed of Composed of several decoding blocks, the spatial resolution is gradually restored through deconvolution layers after the input encoded features are processed, mapping the low-dimensional spatiotemporal features back to the original rainfall field size, thus ensuring high-resolution output of the predicted image. ； 。 7. The intelligent typhoon rainfall prediction method based on multi-scale spatiotemporal coupling according to claim 6, characterized in that, In step 5), the time decoder uses a ConvNormReLU unit to extend the time channel; Features obtained from the spatial decoder The time dimension and channel dimension are concatenated to form a tensor of size T×C, which is then mapped to K×C, where K is the target prediction length. Subsequently, the resulting feature map is subjected to dimensionality transformation to obtain the target prediction dimension. ; The calculation process is as follows: 。 8. The intelligent typhoon rainfall prediction method based on multi-scale spatiotemporal coupling according to claim 1, characterized in that, The loss function is used to calculate the typhoon core region separately. , peripheral area and the entire region Precipitation forecast error; set up These represent the predicted and actual precipitation fields, respectively. A binary mask for the cyclone's core region. It serves as a mask for the outer area. The regional MSE loss function is defined as follows: ； Total loss function: ； in, This represents the actual precipitation of the b-th sample at time t. A binary mask for the kernel region. For the outer area; , and These are the weighting coefficients.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 8.

Images