Short-term and imminent precipitation forecasting method, system, medium, and device based on spatio-temporal decoupling

Abstract

The present invention discloses a short-term and imminent precipitation forecasting method, system, medium, and device based on spatio-temporal decoupling, which relates to the technical field of short-term and imminent precipitation forecasting. The method includes: obtaining a short-term and imminent precipitation forecasting data set; constructing a short-term and imminent precipitation forecasting model, including an encoder part, a time predictor, and a decoder part. The encoder part includes a plurality of encoders and decoders with the same quantity. The encoder part, the time predictor, and the decoder part are connected in series in sequence to form a U-shaped structure. The plurality of encoders are connected in series through downsampling, and the plurality of decoders are connected in series. The result of each encoder is connected to the corresponding decoder through a skip connection; training and testing the model to obtain a final short-term and imminent precipitation forecasting model, and inputting the short-term and imminent precipitation forecasting data to be measured into the final short-term and imminent precipitation forecasting model to obtain a short-term and imminent precipitation forecasting result. The present invention can achieve accurate short-term and imminent precipitation forecasting.

Description

Technical Field

[0001] The present invention relates to the technical field of short-term and imminent precipitation forecasting, and particularly to a short-term and imminent precipitation forecasting method, system, medium, and device based on spatio-temporal decoupling. Background Art

[0002] Rainfall, as one of the most common meteorological phenomena in daily life, has a profound impact on all aspects of life and is also one of the important causes of many natural disasters. Accurate and timely short-term and imminent precipitation forecasting (i.e., high-time-resolution meteorological forecasting within a short time range, generally within 0-2 hours) is crucial and greatly affects human production and life, such as assisting flood warnings, traffic control, crop harvesting, and weather warnings for offshore fishing. However, different from large-scale weather systems, rainfall has the characteristics of strong suddenness, complex and diverse mechanisms, and short life cycles, and short-term and imminent precipitation forecasting has always been a research difficulty.

[0003] Traditional methods for short-term and imminent precipitation forecasting mainly include: numerical weather prediction method and extrapolation method based on radar echo. The former's complex calculations usually rely on high-performance computers, with high computing power costs, and are usually greatly affected by the initial condition field, resulting in poor forecasting within the 0-2h lead time and difficulty in meeting the requirements of accuracy and timeliness in refined forecasting. The latter, although having high timeliness and relatively clear prediction images, also has limitations. Its extrapolation forecasting ability will rapidly decline as the forecasting time limit extends, with obvious deficiencies in forecasting ability and obvious distortion effects in the prediction map, unable to meet the needs of practical applications.

[0004] Existing deep learning models are mainly divided into two types: based on autoregressive methods and based on non-autoregressive methods. However, the former has problems such as difficult training convergence and easy occurrence of gradient disappearance, while the latter lacks the ability to extract time features. Summary of the Invention

[0005] The purpose of the present invention is to: in order to solve the problem of poor time feature extraction ability in existing short-term and imminent precipitation forecasting methods, a short-term and imminent precipitation forecasting method based on spatio-temporal decoupling is proposed, including the following steps:

[0006] S1. Obtain a short-term and imminent precipitation forecasting dataset, perform preprocessing, and divide the preprocessed dataset into a training set and a test set;

[0007] S2. Construct a short-term and imminent precipitation forecasting model, including an encoder part, a time predictor, and a decoder part. The encoder part includes multiple encoders, and the number of decoders in the decoder part is the same as the number of encoders;

[0008] The encoder part, the time predictor, and the decoder part are connected in series in sequence to form a U-shaped structure. Multiple encoders are connected in series through downsampling, multiple decoders are connected in series, and the result of each encoder is connected to the corresponding decoder through a skip connection;

[0009] The feature map generated by the last encoder in the encoder part is input into the time predictor, and the feature map output by the time predictor is input into the decoder part, and the decoder part obtains the prediction of the precipitation situation at future times;

[0010] S3. Use the training set and the test set to train and test the short-term and nowcasting precipitation prediction model to obtain the final short-term and nowcasting precipitation prediction model, and input the short-term and nowcasting precipitation prediction data to be measured into the final short-term and nowcasting precipitation prediction model to obtain the short-term and nowcasting precipitation prediction result.

[0011] Furthermore, the encoder includes: a multi-scale self-attention mechanism module, a multiplicative convolution module, and an adaptive dynamic weighting module;

[0012] The multi-scale self-attention mechanism module is used to extract the global spatial features of the short-term and nowcasting precipitation prediction data;

[0013] The multiplicative convolution module is used to extract the local spatial features of the short-term and nowcasting precipitation prediction data;

[0014] The adaptive dynamic weighting module is used to fuse the global and local spatial features.

[0015] Furthermore, the multi-scale self-attention mechanism module includes two layer normalization modules, one downsampling module, three attention mechanisms, one upsampling module, one convolutional layer, and one feed-forward neural network;

[0016] The input feature map of the multi-scale self-attention mechanism module passes through the first layer normalization module and the downsampling module to obtain feature maps of three different scales: large, medium, and small. The feature maps of the three different scales respectively pass through the three attention mechanisms to obtain attention feature maps of the three different scales: large, medium, and small. The attention feature maps of the small and medium scales are upsampled and then concatenated with the attention feature map of the large scale to obtain the concatenated attention feature map. The concatenated attention feature map passes through the convolutional layer, the second layer normalization module, and the feed-forward neural network in sequence. Among them, the input feature map of the multi-scale self-attention mechanism module is connected to the output of the convolutional layer through a skip connection, and the input of the second layer normalization module is connected to the output of the feed-forward neural network through a skip connection. Finally, the global spatial features output by the multi-scale self-attention mechanism module are obtained.

[0017] Furthermore, the multiplication convolution module includes: a layer normalization module and two depthwise separable convolution layers; after the input of the multiplication convolution module is processed by the layer normalization module, it is input into the first depthwise separable convolution layer. The output feature map of the first depthwise separable convolution layer is evenly divided into two branches, and the result of multiplying the feature maps of the two branches is passed through the second depthwise separable convolution layer, where the output of the layer normalization module is skip-connected to the output of the second depthwise separable convolution layer, and finally the local spatial features output by the multiplication convolution module are obtained.

[0018] Furthermore, the adaptive dynamic weighting module includes three convolution layers and a sigmoid layer; among them, the global spatial feature x and the local spatial feature y pass through the first convolution layer and the second convolution layer respectively to generate two feature vectors and , the is input into the sigmoid layer to obtain the dynamic weight α, and the result of is passed through the third convolution layer. The global spatial feature x and the local spatial feature y are skip-connected to the output of the third convolution layer, and finally the fused feature that fuses the global and local spatial features output by the adaptive dynamic weighting module is obtained.

[0019] Furthermore, the time predictor includes: two reshape layers and a prediction module with the architecture of the encoder block in the Transformer. The prediction module includes N prediction units, and each prediction unit includes: a Fourier self-attention mechanism layer, two layer normalization layers, and a feed-forward neural network;

[0020] The input of the time predictor passes through the first reshape layer, the prediction module, and the second reshape layer in sequence to obtain the output of the time predictor.

[0021] Furthermore, the loss function during the training of the model is as follows:

[0022]

[0023]

[0024] Among them, represents the weighted loss function, F represents the number of frames of the image sequence, H represents the height of the feature map, W represents the width of the feature map, represents the weight of the feature map with the number of frames f, height h, and width w, represents the prediction result of the feature map with the number of frames f, height h, and width w, represents the true label value of the feature map with the number of frames f, height h, and width w, and x represents the pixel value of the real image point.

[0025] The present invention also provides a short-term and impending precipitation forecasting system based on spatio-temporal decoupling, comprising:

[0026] A data acquisition module, configured to acquire a short-term and impending precipitation forecasting data set, perform preprocessing on the data set, and divide the preprocessed data set into a training set and a test set;

[0027] A model construction module, configured to construct a short-term and impending precipitation forecasting model, including an encoder part, a time predictor, and a decoder part. The encoder part includes multiple encoders, and the number of decoders in the decoder part is the same as the number of encoders;

[0028] The encoder part, the time predictor, and the decoder part are connected in series in sequence to form a U-shaped structure. The multiple encoders are connected in series through downsampling, and the multiple decoders are connected in series. The result of each encoder is connected to the corresponding decoder through a skip connection;

[0029] The feature map generated by the last encoder in the encoder part is input into the time predictor, and the feature map output by the time predictor is input into the decoder part, and the decoder part obtains the prediction of the precipitation situation at a future time;

[0030] A result acquisition module, configured to train and test the short-term and impending precipitation forecasting model using the training set and the test set to obtain a final short-term and impending precipitation forecasting model, and input the short-term and impending precipitation forecasting data to be measured into the final short-term and impending precipitation forecasting model to obtain a short-term and impending precipitation forecasting result.

[0031] The present invention also provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the above-mentioned short-term and impending precipitation forecasting method based on spatio-temporal decoupling is implemented.

[0032] The present invention also provides an electronic device, including a processor and a memory, the processor is connected to the memory, wherein the memory is used to store a computer program, the computer program includes computer-readable instructions, and the processor is configured to call the computer-readable instructions to execute the above-mentioned short-term and impending precipitation forecasting method based on spatio-temporal decoupling.

[0033] The beneficial effects brought by the technical solution provided by the present invention are:

[0034] Aiming at the long-distance dependence problem in short-term and impending precipitation forecasting, the present invention constructs a multi-scale self-attention module to extract global spatial features, constructs a multiplicative convolution module to extract local spatial features, constructs an adaptive dynamic weighting module to guide the model to adaptively fuse spatial features, and constructs a time predictor based on the Fourier self-attention mechanism to extract time features, so as to achieve accurate short-term and impending precipitation forecasting. Description of the Drawings

[0035] Figure 1It is the flowchart of the short-term and imminent precipitation forecasting method based on spatio-temporal decoupling in the embodiments of the present invention;

[0036] Figure 2 It is the structural diagram of the short-term and imminent precipitation forecasting model in the embodiments of the present invention;

[0037] Figure 3 It is the structural diagram of the multi-scale self-attention mechanism module in the embodiments of the present invention;

[0038] Figure 4 It is the structural diagram of the multiplication convolution module in the embodiments of the present invention;

[0039] Figure 5 It is the structural diagram of the adaptive dynamic weighting module in the embodiments of the present invention;

[0040] Figure 6 It is the schematic diagram of the time predictor module in the embodiments of the present invention;

[0041] Figure 7 It is the visual comparison between the model in the embodiments of the present invention and six other comparison models;

[0042] Figure 8 It is the comparison chart of CSI and HSS scores of the model in the embodiments of the present invention and six other comparison models at different time steps. The CSI score is shown on the left, and the HSS score is shown on the right. Among them, Figure 8 the CSI and HSS thresholds in (a) are 0.5 mm / h, Figure 8 the CSI and HSS thresholds in (b) are 2 mm / h, Figure 8 the CSI and HSS thresholds in (c) are 5 mm / h, Figure 8 the CSI and HSS thresholds in (d) are 10 mm / h, Figure 8 the CSI and HSS thresholds in (e) are 30 mm / h;

[0043] Figure 9 It is the block diagram of an electronic device in an exemplary embodiment of the embodiments of the present invention. Detailed implementation manners

[0044] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below in conjunction with the accompanying drawings.

[0045] Embodiment 1: The flowchart of the short-term and imminent precipitation forecasting method based on spatio-temporal decoupling in the embodiments of the present invention is as Figure 1 , and specifically includes the following steps:

[0046] S1. Obtain the short-term and imminent precipitation forecasting dataset, perform preprocessing, and divide the preprocessed dataset into a training set and a test set.

[0047] Specifically, a precipitation dataset is generated using two C-band Doppler weather radar stations in the study area. The two radars perform four 360° azimuth scans at beam elevation angles of 0.3°, 1.1°, 2.0°, and 3.0° around the vertical axis. A total of 420,000 rainfall maps at 5-minute intervals for the four years from 2016 to 2019 are included.

[0048] Due to the high imbalance between rainy and non-rainy data in the images, images with at least 50% of pixels containing any rainfall amount are selected as the training set and the validation set. The dataset is divided into a training set (from 2016 to 2018) and a test set (in 2019). The original rainfall maps are sized 765×700, and one pixel corresponds to the cumulative rainfall in one square kilometer for the most recent five minutes. The rainfall amount is represented as an integer value in hundredths of a millimeter. For example, the value 12 indicates a rainfall amount of 0.12 millimeters in the past 5 minutes. In the present invention, the values of the training set and the test set are divided by the highest value that appears in the training set to standardize the data. The calculation formula is as follows, where is the standardized data, x is the original data, and max is the highest value among all the data.

[0049]

[0050] Meanwhile, considering that some pixel points in the original images have no numerical values, which can easily lead to difficulties in neural network training, the original images are cropped to a size of 288×288. Since this dataset contains images with very little or no rainfall, in order to prevent the network from being biased towards predicting zero values, image screening is performed to ensure that at least 50% of the pixels in the images are rain pixels.

[0051] S2. Construct a short-term precipitation forecasting model, including an encoder part, a time predictor, and a decoder part. The encoder part includes multiple encoders; the decoder in the decoder part corresponds one-to-one with the encoder, and the number is the same as the number of encoders.

[0052] The encoder part, the time predictor, and the decoder part are connected in series in sequence to form a U-shaped structure. The encoders and decoders under the U-shaped structure correspond one-to-one. Multiple encoders are connected in series through downsampling, and multiple decoders are connected in series. The result of each encoder is connected to the corresponding decoder through a skip connection.

[0053] The feature map generated by the last encoder in the encoder part is input into the time predictor, and the feature map output by the time predictor is input into the decoder part, and the decoder part obtains the prediction of the precipitation situation at future times.

[0054] In a preferred embodiment, the short-term and imminent precipitation forecasting model includes three encoders and three decoders. After preprocessing the obtained short-term and imminent precipitation forecasting data set, through a patch embedding module, the original two-dimensional image is converted into a series of one-dimensional patch embeddings to obtain the embedded representation of the short-term and imminent precipitation forecasting data. The embedded representation is input into the first decoder to obtain the first feature map. The first feature map is input into the second encoder after passing through the first downsampling module to obtain the second feature map. The second feature map is input into the third encoder after passing through the second downsampling module to obtain the third feature map. The third feature map passes through the third downsampling module to obtain the fourth feature map; the fourth feature map is input into the time predictor to obtain the fifth feature map. For the structural diagram of the short-term and imminent precipitation forecasting model in the embodiment of the present invention, refer to Figure 2 。

[0055] The structure of the decoder is an upsampling module, a decoding unit, and a convolutional layer connected in series in sequence. The first decoder, the second decoder, and the third decoder are connected in series in sequence. The fifth feature map is input into the upsampling module of the first decoder. The third feature map obtained by the third encoder is connected to the output of the decoding unit of the first decoder through a skip connection. After the third feature map and the output of the decoding unit of the first decoder are concatenated, they are input into the convolutional layer of the first decoder. The second feature map obtained by the second encoder is connected to the output of the decoding unit of the second decoder through a skip connection. After the second feature map and the output of the decoding unit of the second decoder are concatenated, they are input into the convolutional layer of the second decoder. The first feature map obtained by the first encoder is connected to the output of the decoding unit of the third decoder through a skip connection. After the first feature map and the output of the decoding unit of the third decoder are concatenated, they are input into the convolutional layer of the third decoder. The output of the third decoder passes through a convolutional layer to obtain the output of the decoder part.

[0056] The encoder includes: a multi-scale self-attention mechanism module, a multiplicative convolution module, and an adaptive dynamic weighting module.

[0057] Considering the characteristics of short-term precipitation forecasting, a larger receptive field is beneficial for capturing and understanding the global changes in precipitation information and identifying the mutual influence of precipitation in distant regions. A multi-scale self-attention mechanism module is constructed. The multi-scale self-attention mechanism module is used to extract the global spatial features of the short-term and imminent precipitation forecasting data and includes self-attention feature extraction units at three scales.

[0058] The multi-scale self-attention mechanism module includes two layer normalization modules, one downsampling module, three attention mechanisms, one upsampling module, one convolutional layer, and one feed-forward neural network. The input feature map of the multi-scale self-attention mechanism module passes through the first layer normalization module and the downsampling module to obtain three feature maps of different scales: large, medium, and small. The three feature maps of different scales respectively pass through three attention mechanisms to obtain three attention feature maps of different scales: large, medium, and small. The small and medium scale attention feature maps are upsampled and then concatenated with the large scale attention feature map to obtain a concatenated attention feature map. The concatenated attention feature map sequentially passes through the convolutional layer, the second layer normalization module, and the feed-forward neural network. Among them, the input feature map of the multi-scale self-attention mechanism module is connected to the output of the convolutional layer through a skip connection, and the input of the second layer normalization module is connected to the output of the feed-forward neural network through a skip connection. Finally, the global spatial features output by the multi-scale self-attention mechanism module are obtained. The multi-scale self-attention mechanism module can simultaneously extract and fuse three precipitation feature maps of different scales, thereby enhancing the network's prediction ability for precipitation of different scales. This module calculates the attention for the entire feature map to evaluate the relationship between different data points. Specifically, the attention mechanism for the large-scale feature map enables the network to analyze the overall trend of precipitation changes in the entire region and capture the global precipitation pattern. The attention mechanism for the medium-scale feature map focuses on the evolution of precipitation events of different intensities and reveals the medium-scale precipitation dynamics. The attention mechanism for the small-scale feature map further refines the understanding of the precipitation details in the local area and reveals the internal connections of precipitation in this area. This hierarchical attention calculation strategy provides a comprehensive perspective for the network, making the short-term precipitation forecast more accurate and detailed. The structural diagram of the multi-scale self-attention mechanism module in the embodiment of the present invention is referred to Figure 3 。

[0059] Although applying the self-attention mechanism on the small-scale feature map can effectively capture the key information in the local area, compared with the convolutional operation, its receptive field is still large, and some information is also lost during the downsampling process. Considering that extreme precipitation events are often confined to a relatively small geographical area, in order to enhance the ability to capture the detailed features of these events, a multiplicative convolution module is constructed. The overall structural schematic diagram of the module is as Figure 4As shown in the figure. The multiplication convolution module is used to extract the local spatial features of short-term precipitation forecast data. The specific operation is as follows: First, the feature tensor is normalized to standardize the data, thus avoiding the problem of gradient vanishing or explosion. Subsequently, the number of channels is expanded to 6 times the input through the depthwise separable convolution layer to enhance the feature representation ability of the model. Given the complexity, variability, and strong nonlinear characteristics of the short-term precipitation forecast task, this module uses multiplication operations to better capture and fit these nonlinear changes. Specifically, the output feature map is evenly divided into two branches. In these two branches, pairwise multiplication of corresponding elements is performed to achieve nonlinear interaction and fusion of features. Finally, through another depthwise separable convolution layer, while enriching the feature information after nonlinear transformation, the number of channels is restored to the original size.

[0060] The multiplication convolution module includes: a layer normalization module and two depthwise separable convolution layers; after the input of the multiplication convolution module is processed by the layer normalization module, it is input into the first depthwise separable convolution layer. The output feature map of the first depthwise separable convolution layer is evenly divided into two branches A and B. The results of pairwise multiplication of the same points at corresponding positions of the feature maps of the two branches A and B (before multiplication, the feature map of A first passes through the ReLU activation function) pass through the second depthwise separable convolution layer, where the output of the layer normalization module is skip-connected to the output of the second depthwise separable convolution layer, and finally the local spatial features output by the multiplication convolution module are obtained.

[0061] It is often difficult to achieve the expected results by simply adding the global spatial features extracted by the multi-scale self-attention mechanism module and the local spatial features extracted by the multiplication convolution module, which may lead to mutual coverage and interference between features. Arbitrarily setting the fixed weights of the two cannot accurately capture the complex dynamic relationships in the data. To solve this problem, an adaptive dynamic weighting module is constructed. The adaptive dynamic weighting module is used to fuse global and local spatial features. The adaptive dynamic weighting module includes three convolution layers and a sigmoid layer; among them, the global spatial feature x and the local spatial feature y pass through the first convolution layer and the second convolution layer respectively to generate two feature vectors and , and is input into the sigmoid layer to obtain the dynamic weight α. The size of the α value reflects the trust degree of the network in the features of the two branches. The larger the α value, the more detailed features the network believes are needed at this time; the smaller the α value, the more global features are needed. The result of passes through the third convolution layer, and the global spatial feature x and the local spatial feature y are skip-connected to the output of the third convolution layer. Finally, the fused feature that fuses global and local spatial features output by the adaptive dynamic weighting module is obtained. The structure diagram of the adaptive dynamic weighting module in the embodiment of the present invention refers to Figure 5 .

[0062] To overcome the limitation that the traditional encoder-decoder architecture cannot fully utilize the time series information when processing precipitation data, a time predictor is constructed. The time predictor includes: two reshape layers and a prediction module with the architecture of the encoder block in the Transformer. The prediction module includes N prediction units, and each prediction unit includes: a Fourier self-attention mechanism layer, two layer normalization layers, and a feed-forward neural network. The input of the time predictor passes through the first reshape layer, the prediction module, and the second reshape layer in sequence to obtain the output of the time predictor. Reshape means changing only the dimensions of a tensor without changing the data itself. Specifically, the time predictor first splits the fourth feature map generated by the encoder part into H×W multivariate time series with a length of C. Each time series represents the past precipitation situation at that point. Based on this historical data, the time predictor predicts the precipitation situation at each point in the future. Considering the significant time features in the precipitation data, the data is transformed into the frequency domain before calculating the self-attention values. The schematic diagram of the time predictor module is as Figure 6 shown.

[0063] S3. Use the training set and the test set to train and test the short-term and nowcasting precipitation prediction model to obtain the final short-term and nowcasting precipitation prediction model. Input the short-term and nowcasting precipitation prediction data to be measured into the final short-term and nowcasting precipitation prediction model to obtain the short-term and nowcasting precipitation prediction results.

[0064] Since the distribution of rainfall amounts in the dataset is extremely unbalanced, a weighted loss function is used. The loss function during model training is as follows:

[0065]

[0066]

[0067] Among them, represents the weighted loss function, F represents the number of frames of the image sequence, H represents the height of the feature map, W represents the width of the feature map, represents the weight of the feature map with the number of frames f, height h, and width w, represents the prediction result of the feature map with the number of frames f, height h, and width w, represents the true label value of the feature map with the number of frames f, height h, and width w, and x represents the pixel value of the true image point.

[0068] Read the training data and convert it into the tensor type, train the model, use the Adam optimizer for training, and set the learning rate to decay by 0.9 times every 4 epochs until the loss value tends to be stable and the model converges.

[0069] The read test data is converted into the tensor type and then fed into the trained model for inference to obtain the inference result. Four metrics, namely weighted root mean square error (BMSE), weighted mean absolute error (BMAE), critical success index (CSI), and Heidke skill score (HSS), are used to evaluate the model performance, and the optimal model is saved.

[0070] The method of the present invention is compared with several models, namely Unet, Rainformer, LPT-QPN, SimVP, Predrnn, and MotionRNN. Table 1 quantitatively compares the prediction results of all models. "R≥t" represents the nowcasting skill score at the threshold t (mm / h). The ranges of CSI and HSS are [0,1], and the higher the value, the better; the ranges of BMSE and BMAE are [0, +∞], and the lower the value, the better.

[0071] Among them, the Unet model is from the paper: U-Net: Convolutional Networks for Biomedical Image Segmentation. UNet is a symmetric network structure, with downsampling on the left side and upsampling on the right side. A series of downsampling operations on the left side can be called the encoder according to their functions, and a series of upsampling operations on the right side can be called the decoder. Rainformer is from the paper: Features Extraction Balanced Network for Radar-Based Precipitation Nowcasting, a feature extraction balanced network for radar-based precipitation nowcasting. LPT-QPN is from the paper: A Lightweight Physics-Informed Transformer for Quantitative Precipitation Nowcasting, a lightweight physics-informed transformer for quantitative precipitation nowcasting. SimVP is from the paper: Simpler yet Better Video Prediction. The SimVP model stacks Inception units in the Translator to form an encoder-decoder structure. This encoder-decoder structure has a total depth of 8 layers, and time features are extracted in this way to obtain the time evolution effect. Predrnn is a recurrent neural network for spatiotemporal sequence prediction, from the paper: Recurrent Neural Networks for Predictive Learning using Spatiotemporal LSTMs. MotionRNN is from the paper: A Flexible Model for Video Prediction with Spacetime-Varying Motions. MotionRNN is an RNN-based video prediction model that can combine multiple prediction models to uniformly model motion trends and instantaneous changes.

[0072] For the convenience of comparison, the results with r > 10 are multiplied by 10, and the results with r > 30 are multiplied by 100. It can be seen that the CSI and HSS indicators of the method of the present invention rank among the top in all threshold cases, which fully proves its superior precipitation forecasting ability. At the same time, the BMSE and BMAE indicators are also the best, indicating that its prediction results have the highest similarity with the actual values. This benefits from the efficient cooperation among the components of the model of the present invention.

[0073] Table 1

[0074]

[0075] Figure 7 is a visual comparison between the model of the embodiment of the present invention and six other comparative models, Figure 7 which intuitively shows the prediction visualization results of all models on the KNMI dataset. Visually, all models have weak prediction capabilities in the heavy precipitation area with a later time frame. Especially after more than 25 minutes, the prediction capabilities of all models almost drop to 0. However, from the quantitative results, it can be seen that the prediction accuracy of TPDTC-Net is significantly higher than that of other models, fully demonstrating its excellent prediction capabilities in dealing with heavy precipitation situations.

[0076] Figure 8 is a comparison chart of CSI and HSS scores of the model of the embodiment of the present invention and six other comparative models at different time steps. The CSI score is shown on the left and the HSS score is shown on the right, where, Figure 8 the CSI and HSS thresholds in (a) are 0.5 mm / h, Figure 8 the CSI and HSS thresholds in (b) are 2 mm / h, Figure 8 the CSI and HSS thresholds in (c) are 5 mm / h, Figure 8 the CSI and HSS thresholds in (d) are 10 mm / h, Figure 8 the CSI and HSS thresholds in (e) are 30 mm / h. Through the Figure 8 comparison, it can be found that the model of the present invention is superior to other models in almost all time steps. This benefits from the efficient utilization of time series information by its time predictor, as well as the effective extraction and fusion of global and local spatial features. Although LPT_QPN performs well, its centertransformer module only focuses on enhancing the model's simulation ability for the non-linear changes of precipitation events, but ignores the dynamic changes of time information, which is one of its main defects. As a video frame prediction model completely based on CNN, SimVP learns time evolution information through its translator, but there are essential differences compared with the time predictor proposed in this paper. SimVP fails to fully consider the combined characteristics of the unique time information and spatial position in precipitation data, so the effect is still inferior to the model of the present invention.

[0077] Embodiment 2: The present invention also proposes a short-term and nowcasting precipitation forecasting system based on spatio-temporal decoupling, including:

[0078] A data acquisition module, configured to acquire a short-term and nowcasting precipitation forecasting dataset, perform preprocessing, and divide the preprocessed dataset into a training set and a test set;

[0079] A model construction module for constructing a short-term and imminent precipitation forecasting model, including an encoder part, a time predictor, and a decoder part. The encoder part includes multiple encoders, and the number of decoders in the decoder part is the same as the number of encoders.

[0080] The encoder part, the time predictor, and the decoder part are connected in series in sequence to form a U-shaped structure. Multiple encoders are connected in series through downsampling, and multiple decoders are connected in series. The result of each encoder is connected to the corresponding decoder through a skip connection.

[0081] The feature map generated by the last encoder in the encoder part is input into the time predictor, and the feature map output by the time predictor is input into the decoder part, and the decoder part obtains the prediction of the precipitation situation at future times.

[0082] A result acquisition module for training and testing the short-term and imminent precipitation forecasting model using a training set and a test set to obtain a final short-term and imminent precipitation forecasting model, and inputting the short-term and imminent precipitation forecasting data to be measured into the final short-term and imminent precipitation forecasting model to obtain the short-term and imminent precipitation forecasting result.

[0083] Example 3: In an exemplary embodiment, it includes a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, it implements the above-mentioned short-term and imminent precipitation forecasting method based on spatio-temporal decoupling.

[0084] Example 4: Please refer to Figure 9 , in an exemplary embodiment, it further includes an electronic device, including at least one processor, at least one memory, and at least one communication bus.

[0085] Among them, a computer program is stored on the memory, and the computer program includes computer-readable instructions. The processor calls the computer-readable instructions stored in the memory through the communication bus to execute the above-mentioned short-term and imminent precipitation forecasting method based on spatio-temporal decoupling.

[0086] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A short-term and imminent precipitation forecasting method based on spatio-temporal decoupling, characterized in that, It includes the following steps: S1. Obtain a short-term and impending precipitation forecast dataset, perform preprocessing, and divide the preprocessed dataset into a training set and a test set; S2. Construct a short-term and impending precipitation forecast model, including an encoder part, a time predictor, and a decoder part. The encoder part includes multiple encoders, and the number of decoders in the decoder part is the same as the number of encoders; The encoder part, the time predictor, and the decoder part are sequentially connected in series to form a U-shaped structure. Multiple encoders are connected in series through downsampling, multiple decoders are connected in series, and the result of each encoder is connected to the corresponding decoder through a skip connection; The feature map generated by the last encoder in the encoder part is input into the time predictor, and the feature map output by the time predictor is input into the decoder part, and the decoder part obtains the precipitation situation prediction for future time; The encoder includes: a multi-scale self-attention mechanism module, a multiplicative convolution module, and an adaptive dynamic weighting module; The multi-scale self-attention mechanism module is used to extract the global spatial features of short-term and impending precipitation forecast data; The multiplicative convolution module is used to extract the local spatial features of short-term and impending precipitation forecast data; The adaptive dynamic weighting module is used to fuse the global and local spatial features; The time predictor includes: two reshape layers and a prediction module with the architecture of an encoder block in a Transformer. The prediction module includes N prediction units, and each prediction unit includes: a Fourier self-attention mechanism layer, two layer normalization layers, and a feed-forward neural network; The input of the time predictor passes through the first reshape layer, the prediction module, and the second reshape layer in sequence to obtain the output of the time predictor; S3. Use the training set and the test set to train and test the short-term and impending precipitation forecast model to obtain the final short-term and impending precipitation forecast model, and input the short-term and impending precipitation forecast data to be measured into the final short-term and impending precipitation forecast model to obtain the short-term and impending precipitation forecast result.

2. The short-term and imminent precipitation forecasting method based on spatio-temporal decoupling according to claim 1, wherein The multi-scale self-attention mechanism module includes two layer normalization modules, a downsampling module, three attention mechanisms, an upsampling module, a convolutional layer, and a feed-forward neural network; The input feature map of the multi-scale self-attention mechanism module passes through the first layer normalization module and obtains three feature maps with different scales through the downsampling module. The three feature maps with different scales pass through three attention mechanisms respectively to obtain three attention feature maps with different scales. The attention feature maps of the two smaller scales among the three scales are upsampled and then spliced with the attention feature map of the largest scale among the three scales to obtain the spliced attention feature map. The spliced attention feature map passes through the convolutional layer, the second layer normalization module, and the feed-forward neural network in sequence. Among them, the input feature map of the multi-scale self-attention mechanism module is connected to the output of the convolutional layer through a skip connection, and the input of the second layer normalization module is connected to the output of the feed-forward neural network through a skip connection, and finally the global spatial features output by the multi-scale self-attention mechanism module are obtained.

3. A short-term and imminent precipitation forecasting method based on spatio-temporal decoupling according to claim 1, characterized in that, The multiplication convolution module includes: a layer normalization module and two depthwise separable convolution layers; after the input of the multiplication convolution module is processed by the layer normalization module, it is input into the first depthwise separable convolution layer, the output feature map of the first depthwise separable convolution layer is evenly divided into two branches, and the result of multiplying the feature maps of the two branches is passed through the second depthwise separable convolution layer, where the output of the layer normalization module is skip-connected to the output of the second depthwise separable convolution layer, and finally the local spatial features output by the multiplication convolution module are obtained.

4. A short-term and imminent precipitation forecasting method based on spatio-temporal decoupling according to claim 1, characterized in that The adaptive dynamic weighting module includes three convolutional layers and one sigmoid layer. Among them, the global spatial feature x and the local spatial feature y pass through the first convolutional layer and the second convolutional layer respectively to generate two feature vectors and , input into the sigmoid layer to obtain the dynamic weight α, and pass the result of through the third convolutional layer. The global spatial feature x and the local spatial feature y are skip-connected to the output of the third convolutional layer, and finally the fused feature that combines the global and local spatial features output by the adaptive dynamic weighting module is obtained.

5. A short-term and imminent precipitation forecasting method based on spatio-temporal decoupling according to claim 1, characterized in that, The loss function during the training of the model is as follows: Among them, represents the weighted loss function, F represents the number of frames of the image sequence, H represents the height of the feature map, and W represents the width of the feature map. represents the weight of the feature map with frame number f, height h, and width w. represents the prediction result of the feature map with frame number f, height h, and width w. represents the true label value of the feature map with frame number f, height h, and width w, and x represents the pixel value of the true image point.

6. A short-term and imminent precipitation forecasting system based on spatio-temporal decoupling, characterized in that, It includes: A data acquisition module, which is used to acquire a short-term precipitation forecast dataset, perform preprocessing, and divide the preprocessed dataset into a training set and a test set; A model construction module, which is used to construct a short-term precipitation forecast model, including an encoder part, a time predictor, and a decoder part. The encoder part includes multiple encoders, and the number of decoders in the decoder part is the same as the number of encoders; The encoder part, the time predictor, and the decoder part are connected in series in sequence to form a U-shaped structure. Multiple encoders are connected in series through downsampling, and multiple decoders are connected in series. The result of each encoder is skip-connected to the corresponding decoder; The feature map generated by the last encoder in the encoder part is input into the time predictor, and the feature map output by the time predictor is input into the decoder part, and the decoder part obtains the prediction of the precipitation situation at future times; The encoder includes: a multi-scale self-attention mechanism module, a multiplication convolution module, and an adaptive dynamic weighting module; The multi-scale self-attention mechanism module is used to extract the global spatial features of the short-term precipitation forecast data; The multiplication convolution module is used to extract the local spatial features of the short-term precipitation forecast data; The adaptive dynamic weighting module is used to fuse the global and local spatial features; The time predictor includes: two reshape layers and a prediction module with the architecture of an encoder block in a transformer. The prediction module includes N prediction units, and each prediction unit includes: a Fourier self-attention mechanism layer, two layer normalization layers, and a feed-forward neural network; The input of the time predictor passes through the first reshape layer, the prediction module, and the second reshape layer in sequence to obtain the output of the time predictor; A result acquisition module, which is used to train and test the short-term precipitation forecast model using the training set and the test set to obtain the final short-term precipitation forecast model, and input the short-term precipitation forecast data to be measured into the final short-term precipitation forecast model to obtain the short-term precipitation forecast result.

7. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the method according to any one of claims 1-5.

8. An electronic device, characterized in that, It includes a processor and a memory, and the processor is interconnected with the memory. Among them, the memory is used to store a computer program, the computer program includes computer-readable instructions, and the processor is configured to call the computer-readable instructions to execute the method according to any one of claims 1-5.

Images