A precipitation prediction method based on deep bibranch structure
By combining the PhyCell physical branch and the SimVP residual branch with a deep bi-branch structure model, the problem of low precipitation prediction accuracy is solved, and more efficient precipitation prediction is achieved.
Patent Information
- Application Number
- CN202410087518.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-22
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2044-01-22
AI Technical Summary
Existing precipitation forecasting methods have low accuracy and struggle to effectively utilize the physical laws and spatiotemporal characteristics of meteorological data.
A deep dual-branch structure model is adopted, combining PhyCell physical branch units and SimVP residual branch units. The radar echo data is processed through encoding, decoding and summing units to extract physical features and residual features. The prediction process is optimized by using CNN and partial differential equations.
It improves the accuracy of precipitation forecasting, and can better utilize physical knowledge and spatiotemporal characteristics to achieve more efficient precipitation forecasting.
Smart Images

Figure CN117908164B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of deep learning and video prediction technology, and in particular to a precipitation prediction method, apparatus, computer-readable storage medium, and computer program product based on a deep dual-branch structure. Background Technology
[0002] Precipitation forecasting research plays a crucial guiding role in agricultural work. In recent years, global warming has led to more frequent extreme precipitation events, especially during the flood season, which seriously threaten people's lives and property. Precipitation is a complex nonlinear process, and its occurrence is subject to highly complex meteorological conditions, as well as various factors such as topography and altitude. Therefore, predicting precipitation is very challenging, and improving forecast accuracy has always been a focus of researchers.
[0003] Currently, with the rapid development of artificial intelligence technology, deep learning algorithms have been applied by many researchers in the field of precipitation forecasting due to their ability to efficiently process large amounts of data feature information, learn the interaction laws between meteorological elements, and accurately describe the nonlinear change process of precipitation.
[0004] On the other hand, with the widespread adoption of smart mobile devices, autonomous vehicles, and video surveillance systems, spatiotemporal data, represented by video, has experienced explosive growth. Spatiotemporal data processing has become a research hotspot in recent years, and deep learning methods have achieved remarkable success in this field. Because meteorological data describing atmospheric motion also possesses spatiotemporal correlation, spatiotemporal prediction networks derived from these works have gradually been used as a purely data-driven method for predicting weather conditions in advance. They employ convolutional layers for spatial feature extraction and then input them into recurrent neural networks to learn the evolutionary trends in the temporal dimension. However, unlike video, the evolution of meteorological data is strictly constrained by physical laws, which are typically ignored by CNNs. In radar echo images, the shape and outline of clouds change as their position moves. To overcome this shortcoming, some researchers have introduced variants of recurrent neural networks to represent partial differential equations (PDEs), thereby better utilizing physical knowledge to optimize the learning process and achieving the expected excellent performance. Physical features refer to dynamic changes that conform to physical laws and can be described by PDEs; while residual features contain static detail features, which can be learned through a CNN structure. Although these works utilize physical knowledge, they either assume that the physics behind the data is explicitly known or are limited to fully visible states, which is rare in general spatiotemporal prediction.
[0005] Therefore, how to provide a precipitation forecasting method with higher accuracy has become an urgent technical problem to be solved. Summary of the Invention
[0006] The purpose of this invention is to provide a precipitation prediction method, apparatus, computer-readable storage medium, and computer program product based on a deep bibranch structure, in order to solve the technical problem of low prediction accuracy in existing precipitation prediction methods.
[0007] To achieve the above objectives, according to a first aspect of the present invention, a precipitation prediction method based on a deep bibranch structure is provided. The precipitation prediction method includes the following steps: acquiring raw radar echo data; preprocessing the raw radar echo data to obtain a raw radar echo image sequence; establishing a deep bibranch structure model; training and testing the deep bibranch structure model to obtain a trained deep bibranch structure model; inputting the raw radar echo image sequence into the trained deep bibranch structure model to output a predicted radar echo image sequence; and calculating future precipitation data based on the radar echo intensity of the predicted radar echo image sequence; wherein the deep bibranch structure model includes... The system includes an encoding unit, a PhyCell physical branch unit, a SimVP residual branch unit, a summing unit, and a decoding unit. The encoding unit processes the predicted radar echo image sequence to obtain a feature map. The feature map flows into the PhyCell physical branch unit and the SimVP residual branch unit, respectively. The PhyCell physical branch unit extracts the physical components from the feature map to obtain physical features. The SimVP residual branch unit extracts the residual components from the feature map to obtain residual features. The summing unit sums the physical features and the residual features to obtain the summed features. The decoding unit restores the summed features to the predicted radar echo image sequence.
[0008] Furthermore, the method of inputting the original radar echo image sequence into the trained deep bi-branch structure model and outputting the predicted radar echo image sequence specifically includes: assuming the existence of a latent space H, the physical dynamics and residual factors in the latent state space H are linearly unentangled; the video evolution in the latent state space H is controlled by equation (1): (1) in, and These represent the extracted physical features and residual features, respectively. Physical features are extracted using PhyCell physical branch units, as shown in equation (2): (2); in, This represents taking the partial derivative of the physical dynamics in the hidden state space to extract body features, where t represents the time step and x represents the input sequence image; Then, SimVP residual branch units are used for detail prediction, as shown in equation (3): (3); in, The expression describes the process of extracting residual features by taking the partial derivative of the residual factor in the latent state space, where t represents the time step and x represents the input sequence image. The input x is mapped to the latent state space H through the encoding unit E, and then flows into the PhyCell physical branch unit and the SimVP residual branch unit, respectively. The PhyCell physical branch unit includes a physical predictor that models the latent dynamics and an input assimilation unit that models the interaction between the latent state and the input data. The physical predictor uses the relationship between CNN and partial differential equations to solve the coefficients of the partial differential equations in reverse through a stacked CNN architecture. The input assimilation unit continuously corrects the prediction by applying the difference between the latent state and the input data to achieve the best prediction effect. The SimVP residual branch unit includes an encoder. The encoder consists of four stacked convolutional modules, which are used to extract spatial features. Group normalization is used instead of layer normalization to better suit the small batches of samples used in this study, and it also has stronger generalization ability. The translator is composed of stacked Inception structures. The Inception structure first passes through a 1×1 convolutional kernel, and then performs convolutions with kernel sizes of (3, 5, 7, 11) respectively. Finally, the feature maps are concatenated. The decoder uses four deconvolutions to reconstruct the ground truth frames, thereby realizing the decoding of the translator information. The decoding unit consists of four convSC modules, each of which consists of a convolutional layer, a GroupNorm layer, and a LeakyReLU layer.
[0009] Furthermore, the method for training and testing the deep bibranch structure model to obtain a trained deep bibranch structure model specifically includes: acquiring a large amount of raw radar echo data and establishing a raw radar echo image sequence dataset; dividing the raw radar echo image sequence dataset into a training set and a test set; and using the training set and test set to train and test the deep bibranch structure model to obtain a trained deep bibranch structure model.
[0010] Furthermore, when training the original deep bi-branch structure model using the training set, the mean squared error is used as the loss function to calculate the error between the model's prediction results and the labeled data; the formula for the mean squared error is as follows:
[0011] in, Indicates the actual value. Let n represent the predicted value and n represent the number of samples. The calculated mean squared error is used to optimize the parameters of the prediction model through the backpropagation algorithm until the required number of iterations is reached.
[0012] Furthermore, the precipitation prediction method also includes the following steps: dividing the original radar echo image sequence into a true previous frame image sequence and a true subsequent frame image sequence; inputting the true previous frame image sequence into a trained deep bi-branch structure model to obtain a predicted subsequent frame image sequence; and comparing the predicted subsequent frame image sequence with the true subsequent frame image sequence to test the performance of the trained deep bi-branch structure model.
[0013] Furthermore, the specific method for preprocessing the raw radar echo data to obtain the raw radar echo image sequence includes: parsing and visualizing the raw radar echo data to obtain raw radar echo images; the raw radar echo data consists of 21 layers, and the remaining 17 layers, after removing the top two and bottom two layers, are aggregated by selecting the maximum pixel value layer by layer according to the same position to obtain aggregated radar echo images; the size of each aggregated radar echo image is set to 128×128, the interval between each two aggregated radar echo images is 6 minutes, and 10 consecutive aggregated radar echo images constitute a complete raw radar echo image sequence.
[0014] Furthermore, the method for calculating future precipitation data based on the radar echo intensity of the predicted radar echo image sequence specifically includes: calculating the surface precipitation intensity using formula (5): (5) in, It refers to the intensity of surface precipitation. It is the radar echo intensity. and It is a constant obtained by fitting measured data.
[0015] According to a second aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described method.
[0016] According to a third aspect of the present invention, a computer-readable storage medium is provided having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.
[0017] According to a fourth aspect of the present invention, a computer program product is provided, comprising a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.
[0018] This invention improves the model used for precipitation prediction by designing a new deep bi-branch structure model that combines PhyCell physical branch units and SimVP residual branch units. Experimental verification shows that the deep bi-branch structure model established in this application has a higher accuracy in precipitation prediction compared with models in related technologies.
[0019] Based on the implementation methods provided in the above aspects, this application can be further combined to provide more implementation methods. Attached Figure Description
[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0021] Figure 1 This is an overall flowchart of the present invention; Figure 2 This is a schematic diagram of the precipitation prediction model built according to the present invention; Figure 3 This is a schematic diagram of the Phycell unit structure used in this invention; Figure 4 This is a schematic diagram of the SimVP unit structure used in this invention; Figure 5 The images show the original radar echo image and the aggregated radar echo image. Figure 6 This is a schematic diagram of the original radar echo image sequence; Figure 7 The experimental results and actual values of the deep bi-branch structure model provided by this invention are compared with those of other models. Detailed Implementation
[0022] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. Unless otherwise specified, the techniques used in the embodiments are conventional means well known to those skilled in the art.
[0023] In an optional embodiment of the present invention, a precipitation prediction method based on a deep bibranch structure is provided. The precipitation prediction method includes the following steps: acquiring the original radar echo data to be predicted; preprocessing the original radar echo data to obtain an original radar echo image sequence; establishing a deep bibranch structure model; training and testing the deep bibranch structure model to obtain a trained deep bibranch structure model; inputting the original radar echo image sequence into the trained deep bibranch structure model to output a predicted radar echo image sequence; and calculating the precipitation data for future times based on the radar echo intensity of the predicted radar echo image sequence. The deep bibranch structure model includes... The system comprises an encoding unit, a PhyCell physical branch unit, a SimVP residual branch unit, a summing unit, and a decoding unit. The encoding unit processes the predicted radar echo image sequence to obtain a feature map. The feature map flows into the PhyCell physical branch unit and the SimVP residual branch unit, respectively. The PhyCell physical branch unit extracts the physical components from the feature map to obtain physical features. The SimVP residual branch unit extracts the residual components from the feature map to obtain residual features. The summing unit sums the physical features and the residual features to obtain a summed feature. The decoding unit restores the summed feature to the predicted radar echo image sequence.
[0024] This invention improves the model used for precipitation prediction by designing a new deep bi-branch structure model that combines PhyCell physical branch units and SimVP residual branch units. Experimental verification shows that the deep bi-branch structure model established in this application has a higher accuracy in precipitation prediction compared with models in related technologies.
[0025] Related technologies use the convlstm model for feature extraction of residual components. However, convlstm is a network unit based on RNN, while the right branch of this model, SimVP, is based on CNN. Compared with the radar echo image dataset, SimVP is more suitable for image processing than RNN. CNN is more adapted to the computing mode of GPU, with fewer parameters and shorter training time. Experiments show that this model has excellent performance, and the experimental data exceeds the original model and other prior models.
[0026] Prior to this application, there was no technical solution for applying SimVP to precipitation prediction. This application innovatively selects SimVP as the residual branching unit and combines it with the PhyCell physical branching unit to form a new deep bi-branching structure model, which is more suitable for precipitation prediction and helps to improve the accuracy of precipitation prediction.
[0027] Optionally, the method of inputting the original radar echo image sequence into the trained deep bi-branch structure model and outputting the predicted radar echo image sequence specifically includes: assuming the existence of a latent space H, the physical dynamics and residual factors in the latent state space H are linearly unentangled; the video evolution in the latent state space H is controlled by equation (1): (1) Among them, and These represent the extracted physical features and residual features, respectively. Physical features are extracted using PhyCell physical branch units, as shown in equation (2): (2); in, This represents taking the partial derivative of the physical dynamics in the hidden state space to extract body features, where t represents the time step and x represents the input sequence image; Then, SimVP residual branch units are used for detail prediction, as shown in equation (3): (3); in, The expression describes the process of extracting residual features by taking the partial derivative of the residual factor in the latent state space, where t represents the time step and x represents the input sequence image. The input x is mapped to the latent state space H through the encoding unit E, and then flows into the PhyCell physical branch unit and the SimVP residual branch unit, respectively. The PhyCell physical branch unit includes a physical predictor that models the latent dynamics and an input assimilation unit that models the interaction between the latent state and the input data. The physical predictor uses the relationship between CNN and partial differential equations to solve the coefficients of the partial differential equations in reverse through a stacked CNN architecture. The input assimilation unit continuously corrects the prediction by applying the difference between the latent state and the input data to achieve the best prediction effect. The SimVP residual branch unit includes an encoder. The encoder consists of four stacked convolutional modules, which are used to extract spatial features. Group normalization is used instead of layer normalization to better suit the small batches of samples used in this study, and it also has stronger generalization ability. The translator is composed of stacked Inception structures. The Inception structure first passes through a 1×1 convolutional kernel, and then performs convolutions with kernel sizes of (3, 5, 7, 11) respectively. Finally, the feature maps are concatenated. The decoder uses four deconvolutions to reconstruct the ground truth frames, thereby realizing the decoding of the translator information. The decoding unit consists of four convSC modules, each of which consists of a convolutional layer, a GroupNorm layer, and a LeakyReLU layer.
[0028] Optionally, the method for training and testing the deep bibranch structure model to obtain a trained deep bibranch structure model specifically includes: acquiring a large amount of raw radar echo data to establish a raw radar echo image sequence dataset; dividing the raw radar echo image sequence dataset into a training set and a test set; and using the training set and test set to train and test the deep bibranch structure model to obtain a trained deep bibranch structure model. In this way, through continuous training and testing, the deep bibranch structure model can achieve better prediction results.
[0029] Optionally, the ratio of training set to test set can be 75% and 25%. Other ratios, such as 80% and 20%, can also be selected as needed.
[0030] Optionally, when training the original deep bi-branch structure model using the training set, the mean squared error is used as the loss function to calculate the error between the model's prediction and the labeled data; the formula for the mean squared error is as follows:
[0031] in, Indicates the actual value. Let n represent the predicted value and n represent the number of samples. The calculated mean squared error is used to optimize the parameters of the prediction model through the backpropagation algorithm until the required number of iterations is reached.
[0032] Optionally, the precipitation prediction method further includes the following steps: dividing the original radar echo image sequence into a true preceding frame image sequence and a true following frame image sequence; inputting the true preceding frame image sequence into a trained deep bi-branch structure model to obtain a predicted following frame image sequence; and comparing the predicted following frame image sequence with the true following frame image sequence to test the performance of the trained deep bi-branch structure model. Figure 7 As shown, the proposed model can obtain a predicted subsequent frame image sequence with a high similarity to the actual subsequent frame image sequence; furthermore, the predicted subsequent frame image sequence obtained by the model of this invention is closer to the actual subsequent frame image sequence than that of the comparison model. This verifies that the method of this invention is feasible and has good performance.
[0033] Optionally, the specific method for preprocessing the original radar echo data to obtain the original radar echo image sequence includes: parsing and visualizing the original radar echo data to obtain the original radar echo image; the original radar echo data has 21 layers, and the remaining 17 layers of data, after removing the top two and bottom two layers, are aggregated by selecting the maximum pixel value layer by layer according to the same position to obtain the aggregated radar echo image; the size of each aggregated radar echo image is set to 128×128, the interval between each two aggregated radar echo images is 6 minutes, and 10 consecutive aggregated radar echo images constitute a complete original radar echo image sequence.
[0034] Optionally, in the preprocessing step, the raw radar echo data can be the raw radar echo data to be predicted, or it can be the raw radar echo data used as a training set.
[0035] Optionally, the method for calculating future precipitation data based on the radar echo intensity of the predicted radar echo image sequence specifically includes: calculating the surface precipitation intensity using formula (5): (5) in, It refers to the intensity of surface precipitation. It is the radar echo intensity. and It is a constant obtained by fitting measured data.
[0036] The technical solution of the present invention will be further described below with reference to the accompanying drawings. Specifically, the present invention provides the following technical solution: A precipitation prediction method based on a deep bibranch structure, such as... Figure 1 As shown, the implementation process consists of four steps: Step 1: Data acquisition and preprocessing: data parsing, data visualization, image aggregation, and generation of the original radar echo sequence. Step 2: Design a deep bi-branch structure model. The left and right branches of this model decouple the spatiotemporal sequence data in the hidden state into physical components and residual components, respectively. The physical dynamic features and residual dynamic features obtained from the two branches are then summed, and the summed dynamic features are restored to the predicted data using a decoder. Step 3: 75% and 25% of the radar echo sequence processed in Step 1 are used as the training and test sets, respectively. The deep bi-branch structure model is trained using the training set, and the test set is input into the trained model to output predicted data. This predicted data is compared with the actual values to test the model's performance, thus obtaining the trained deep bi-branch structure model. Step 4: The radar echo data sequence obtained in Step 1 is input into the trained deep bi-branch structure model to obtain the predicted radar echo image sequence. Based on the radar echo intensity of the predicted radar echo image sequence, the future precipitation data is output, thereby achieving precipitation prediction.
[0037] In one specific embodiment of this application, the first step specifically includes: obtaining a raw radar echo image by parsing and visualizing the raw radar echo data from the Northeast Regional Meteorological Center for July and August 2012-2013. The raw radar echo image comprises 21 layers. First, the remaining 17 layers (excluding the top two and bottom two layers) are aggregated by selecting the maximum pixel value layer by layer at the same position, resulting in an aggregated radar echo image "MAX". The raw radar echo image and the aggregated radar echo image are shown below. Figure 5 As shown; next, sequence generation is performed, setting the size of each aggregated radar echo image to a 128×128 grayscale image, with a 6-minute interval between every two radar echo images, and 10 consecutive radar echo images constitute a complete radar echo image sequence, as shown. Figure 6 As shown.
[0038] The second step specifically includes: designing a deep bi-branch structure model, the detailed structure of which is as follows... Figure 2 As shown, the left branch captures the physical dynamic features of the object's motion through the physical component extraction unit, while the right branch captures the remaining residual features from the object's motion through the SimVP network unit, obtaining texture detail features. The specific structure of the PhyCell physics branch unit includes a physics predictor and an input assimilation unit, as shown below. Figure 3 As shown; the specific structure of the SimVP residual branch unit includes: encoder, translator and decoder, as follows. Figure 4 As shown. The specific structure of the decoding unit includes the decoder in the deep dual-branch structure, which consists of four convSC modules, each of which consists of a convolutional layer, a GroupNorm layer, and a LeakyReLU layer.
[0039] The model assumes the existence of a conceptual hidden state space H, where the physical dynamics and residual factors are linearly unentangled. Therefore, the video evolution in the hidden state space H is governed by equation (1): (1) in, and Representing physical characteristics and residual characteristics respectively, the left branch physical characteristics Constrained by the partial differential equation, as shown in equation (2): (2) This represents taking the partial derivative of the physical dynamics in the hidden state space to extract physical features, where t represents the time step and x represents the input sequence image.
[0040] SimVP is then used to perform detail prediction in the hidden state space, as shown in equation (3): (3) The expression represents taking the partial derivative with respect to the residual factor in the latent state space to extract residual features. t represents the time step, and x represents the input sequence image.
[0041] The input x is mapped to the latent state space H by the encoder E, and then flows into the physics branch and the residual branch respectively. In the physics branch, physical dynamic features are extracted by PhyCell units. The PhyCell unit is divided into two parts: a physics predictor that models the latent dynamics and a bias corrector that models the interaction between the latent state and the input data. The physics predictor uses the relationship between CNN and partial differential equations to solve the coefficients of the partial differential equations in reverse through a stacked CNN architecture. The bias corrector continuously corrects the prediction by applying the difference between the latent state and the input data to achieve the best prediction effect. In the residual branch, the SimVP network unit encoder, translator and decoder results are used for prediction. The encoder stacks four convolutional modules. The stacking is used to extract spatial features, and group normalization is used instead of layer normalization to better suit the small batch samples used in this study, and also has stronger generalization ability. The translator is composed of stacked Inception structures. The Inception structure first passes through a 1×1 convolutional kernel, and then uses convolutions with kernel sizes of (3, 5, 7, 11) to perform convolutions. Finally, the feature maps are concatenated. The decoder uses four deconvolutions to reconstruct the ground reality frame, thereby decoding the translator information. The decoder is similar to the encoder, except that the convolution operation is replaced with a deconvolution operation.
[0042] The third step specifically includes: training the constructed deep bi-branch structure model using the partitioned dataset, and using the root mean square error (RMSE) as the loss function during training to calculate the error between the model's prediction results and the labeled data; the formula for RMSE is as follows: (4) in Indicates the actual value. The predicted value is represented by the error, and then the parameters of the prediction model are optimized through the backpropagation algorithm until the number of iterations is reached.
[0043] During testing, Structural Similarity SSIM (SSIM) is added as an evaluation metric to assess the quality of the generated predicted image. The closer the SSIM value is to 1, the higher the quality of the predicted image. The formula is as follows: (5) in Represents the average pixel value of the predicted image. To refer to the average pixel value of the image, Represents the variance of the predicted image, To refer to the variance of the image, Represents the covariance between the predicted image and the reference image. and It is a parameter used to stabilize weak denominators.
[0044] The fourth step specifically includes: inputting the radar echo data sequence obtained in step one into the trained model to obtain the predicted radar echo image sequence.
[0045] To verify that the model provided by this invention can be implemented and has superior performance, the radar echo data sequence obtained in step one is divided into two parts: a real previous frame image sequence and a real subsequent frame image sequence. The real previous frame image sequence is input into the trained deep dual-branch structure model to obtain the predicted subsequent frame image sequence. Figure 7 As shown.
[0046] Based on the radar echo intensity of the predicted radar echo image sequence, the precipitation data for future times is output, thereby realizing precipitation prediction; the surface precipitation intensity is calculated using formula (5): (5) in, This is the ground precipitation value. It is the radar echo intensity. and It is a constant obtained by fitting measured data.
[0047] To verify that the model of this invention outperforms other models, the deep bi-branch structure model provided by this invention is compared with existing models using the moving-MNIST dataset to verify its superiority. The moving-MNIST dataset is an artificially synthesized dataset that requires no preprocessing and is easily obtained; therefore, it is widely used in the field of spatiotemporal prediction. The dataset consists of two moving 28×28 numbers against a 64×64 black background grid. The numbers move independently within the grid and bounce off the boundaries. By assigning different initial positions and velocities to each number, an infinite number of sequences can be obtained. This experiment predicts the next 10 frames from the previous 10 frames. The mean squared error (MSE) and structural similarity (SSIM) of each model are obtained, and the experimental results are shown in Table 1. Table 1 shows that the deep bi-branch structure model provided by this invention has the lowest mean squared error (MSE) and the highest structural similarity (SSIM), indicating that its performance is superior to other control models.
[0048] Table 1
[0049] To further verify that the model of this invention can achieve precipitation prediction and outperform other models, the deep bibranch structure model provided by this invention was compared with existing models using the radar echo sequence processed in step one to verify the model's effectiveness. The mean squared error (MSE) and structural similarity (SSIM) of each model were obtained, and the experimental results are shown in Table 2. As can be seen from Table 2, the deep bibranch structure model provided by this invention has the smallest mean squared error (MSE) and the highest structural similarity (SSIM), demonstrating the effectiveness of this model in precipitation prediction.
[0050] Table 2
[0051] This invention discloses a precipitation prediction method based on a deep bi-branch structure, belonging to the field of meteorological technology. Specifically, it includes: first, data preprocessing; then, designing a deep bi-branch structure model, where the left and right branches decouple the spatiotemporal sequence data in the latent state into physical components and residual components, respectively; summing the physical dynamic features and residual dynamic features obtained from the two branches; and using a decoder to restore the summed dynamic features to the predicted data; training the constructed deep bi-branch structure model using the data; and finally, inputting radar echo data into the trained model to output future precipitation data, thereby achieving precipitation prediction.
[0052] This invention has good regional expansion capability, focuses on the dynamic information of physical constraints, and can produce more accurate prediction results.
[0053] The precipitation prediction method based on a deep bi-branch structure provided by this invention can autonomously learn the spatiotemporal characteristics of precipitation and predict the changing trends of future video frames.
[0054] This invention proposes and evaluates a deep dual-branch architecture model, which, compared with previous deep models, pays more attention to the physical dynamic information during the motion of objects. The left and right branches respectively obtain the physical dynamic features in the hidden state and the residual dynamic features. It combines the tested physical laws with the deep network model to better predict spatiotemporal sequences. Experiments on radar echo datasets have verified that the performance of this model is better than other control models, and it can produce more accurate prediction results. It can generate precipitation prediction data more accurately and optimize existing precipitation prediction results.
[0055] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above method.
[0056] The present invention also provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.
[0057] The present invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.
[0058] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A precipitation prediction method based on a deep bibranch structure, characterized in that, The precipitation prediction method includes the following steps: The raw radar echo data is acquired and preprocessed to obtain the raw radar echo image sequence. Establish a deep bi-branch structure model; train and test the deep bi-branch structure model to obtain a trained deep bi-branch structure model; The original radar echo image sequence is input into the trained deep dual-branch structure model, and the predicted radar echo image sequence is output. Based on the radar echo intensity of the predicted radar echo image sequence, calculate the precipitation data for future times; The deep dual-branch structure model includes an encoding unit, a PhyCell physical branch unit, a SimVP residual branch unit, a summing unit, and a decoding unit. The encoding unit processes the predicted radar echo image sequence to obtain a feature map. The feature map flows into the PhyCell physical branch unit and the SimVP residual branch unit, respectively. The PhyCell physical branch unit extracts the physical components from the feature map to obtain physical features. The SimVP residual branch unit extracts the residual components from the feature map to obtain residual features. The summing unit sums the physical features and the residual features to obtain a summed feature. The decoding unit restores the summed feature to the predicted radar echo image sequence.
2. The precipitation prediction method based on a deep bi-branch structure according to claim 1, characterized in that, The method for inputting the original radar echo image sequence into the trained deep dual-branch structure model and outputting the predicted radar echo image sequence specifically includes: Assume there exists a hidden space H, in which the physical dynamics and residual factors are linearly unentangled; the video evolution in the hidden state space H is controlled by equation (1): (1) in, and These represent the extracted physical features and residual features, respectively. Physical features are extracted using PhyCell physical branch units, as shown in equation (2): (2) This represents taking the partial derivative of the physical dynamics in the hidden state space to extract body features, where t represents the time step and x represents the input sequence image; Then, SimVP residual branch units are used for detail prediction, as shown in equation (3): (3) This indicates that the partial derivative of the residual factor in the latent state space is taken to extract residual features, where t represents the time step and x represents the input sequence image; The input x is mapped to the hidden state space H through the coding unit E, and then flows into the PhyCell physical branch unit and the SimVP residual branch unit respectively. The PhyCell physics branch unit includes a physics predictor that models the latent dynamics and an input assimilation that models the interaction between the latent state and the input data. The physics predictor uses the relationship between CNN and partial differential equations to solve the coefficients of the partial differential equations in reverse through a stacked CNN architecture. The input assimilation continuously corrects the results by applying the difference between the latent state and the input data to achieve the best prediction effect. The SimVP residual branch unit includes an encoder, a translator, and a decoder. The encoder stacks four convolutional modules to extract spatial features and uses group normalization instead of layer normalization to better suit the small batches of samples used in this study, while also having stronger generalization ability. The translator is composed of stacked Inception structures. The Inception structure first uses a 1×1 convolutional kernel, and then uses convolutions with kernel sizes of (3, 5, 7, 11) to perform convolutions respectively, and finally concatenates the feature maps. The decoder uses four deconvolutions to reconstruct the ground truth frames, thereby realizing the decoding of translator information. The decoding unit consists of four layers of convSC modules, each of which consists of a convolutional layer, a GroupNorm layer, and a LeakyReLU layer.
3. The precipitation prediction method based on a deep bi-branch structure according to claim 1, characterized in that, The specific methods for training and testing the deep bi-branch structure model to obtain a trained deep bi-branch structure model include: Acquire a large amount of raw radar echo data and establish a raw radar echo image sequence dataset. The original radar echo image sequence dataset is divided into a training set and a test set; The deep bi-branch structure model is trained and tested using the training set and the test set to obtain a trained deep bi-branch structure model.
4. The precipitation prediction method based on a deep bi-branch structure according to claim 3, characterized in that, When training the original deep bi-branch structure model using the training set, the mean squared error is used as the loss function to calculate the error between the model's prediction results and the labeled data; the formula for the mean squared error is as follows: in, Indicates the actual value. This represents the predicted value, and n represents the sample size. The calculated mean squared error is used to optimize the parameters of the prediction model through backpropagation until the required number of iterations is reached.
5. The precipitation prediction method based on a deep bi-branch structure according to claim 1, characterized in that, The precipitation prediction method also includes the following steps: The original radar echo image sequence is divided into a real first frame image sequence and a real second frame image sequence. The real previous frame image sequence is input into the trained deep dual-branch structure model to obtain the predicted next frame image sequence. The predicted subsequent frame image sequence is compared with the actual subsequent frame image sequence to test the performance of the trained deep dual-branch structure model.
6. The precipitation prediction method based on a deep bi-branch structure according to claim 1, characterized in that, The specific method for preprocessing the raw radar echo data to obtain the raw radar echo image sequence includes: The raw radar echo data is analyzed and visualized to obtain the raw radar echo image; The original radar echo image has 21 layers. The remaining 17 layers of data, after removing the top two layers and the bottom two layers, are aggregated by selecting the maximum pixel value of each layer at the same position to obtain the aggregated radar echo image. Each aggregated radar echo image was sized to 128×128, with a 6-minute interval between each pair of aggregated radar echo images. Ten consecutive aggregated radar echo images constituted a complete sequence of original radar echo images.
7. The precipitation prediction method based on a deep bi-branch structure according to claim 1, characterized in that, The method for calculating future precipitation data based on the radar echo intensity of the predicted radar echo image sequence specifically includes: calculating the surface precipitation intensity using formula (5): (5) in, It refers to the intensity of surface precipitation. It is the radar echo intensity. and It is a constant obtained by fitting measured data.
8. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.
9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 7.
10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 7.
Citation Information
Patent Citations
Weather prediction method based on physical attention space-time prediction network
CN116520458A
