A Physical Constraint-Based Near-Future Rainfall Prediction Method

Abstract

The present invention discloses a method for predicting approaching rainfall based on physical constraints, including: S1, preprocessing each frame of the meteorological radar echo sequence to obtain a corresponding rainfall image sequence; S2, using the first p frames of rainfall images in the rainfall image sequence to predict the next q frames of rainfall images based on a pre-trained advection simulator as the advection prediction result, where p + q = N, and N is the total number of rainfall images in the rainfall image sequence; S3, taking the advection prediction result as a physical constraint and combining it with the first p frames of rainfall images, and inputting them into a pre-trained approaching rainfall prediction network model to obtain the final rainfall prediction result. The approaching rainfall prediction network model includes a physical guidance prediction module and a prediction result refinement module. This method can generate strong rainfall prediction results with smaller errors, and introduces physical constraints in the process, realizing efficient prediction of approaching rainfall.

Description

Technical Field

[0001] The present invention belongs to the field of rainfall prediction, and particularly relates to a nowcasting rainfall prediction method based on physical constraints. Background Art

[0002] Recently, meteorological disasters have occurred frequently in various places. For example, common meteorological phenomena such as extreme precipitation can cause meteorological disasters such as floods and debris flows, resulting in significant social and ecological losses and economic losses. Therefore, the research on rainfall prediction has very important practical significance.

[0003] Currently, meteorological agencies for rainfall prediction mainly use numerical calculation methods. However, since a large amount and variety of meteorological data are required to simulate the initial field for prediction, traditional numerical calculation methods have problems such as difficult data acquisition and large consumption of computing resources. Under this background, deep learning methods have gradually developed. The rainfall prediction method based on deep learning belongs to a data-driven method. Compared with traditional methods, the required data is smaller and the required computing resources are less. However, most models belong to black-box models, only relying on data-driven and ignoring the importance of applying dynamics and physical principles related to rainfall in deep learning. At the same time, the existing deep learning methods have poor prediction effects on high-intensity rainfall areas, and disasters mainly occur in high-intensity rainfall areas. Therefore, the present invention proposes a nowcasting rainfall prediction method based on physical constraints. Summary of the Invention

[0004] The purpose of the present invention is to propose a nowcasting rainfall prediction method based on physical constraints for the above problems, which can efficiently and accurately predict the rainfall area, especially the high-intensity rainfall area.

[0005] To achieve the above purpose, the technical solutions adopted by the present invention are as follows:

[0006] A nowcasting rainfall prediction method based on physical constraints proposed by the present invention includes the following steps:

[0007] S1. Preprocess each frame of image in the meteorological radar echo sequence to obtain a corresponding rainfall image sequence;

[0008] S2. Based on a pre-trained advection simulator, use the first p frames of rainfall images in the rainfall image sequence to predict the next q frames of rainfall images as the advection prediction result, where p + q = N, and N is the total number of rainfall images in the rainfall image sequence;

[0009] S3. Use the advection prediction result as a physical constraint and combine it with the first p frames of rainfall images, and input them into a pre-trained nowcasting rainfall prediction network model to obtain the final rainfall prediction result. The nowcasting rainfall prediction network model includes a physics-guided prediction module and a prediction result refinement module, where:

[0010] The physical guidance prediction module includes an advection prediction result feature extraction module and a gate mechanism module. The advection prediction result feature extraction module is used to separately extract the maximum value and the average value of the advection prediction result through a parallel first calculation module and a second calculation module, and add the maximum value and the average value to obtain a first extracted feature. The first calculation module includes a global max pooling layer and a fully connected layer connected in sequence, and the second calculation module includes a global average pooling layer and a fully connected layer connected in sequence;

[0011] The gate mechanism module is used to splice the previous p-frame rainfall images and the advection prediction result to obtain a spliced feature, convert the spliced feature through a convolutional layer into a second extracted feature with the same dimension as the advection prediction result, then multiply the second extracted feature by the first extracted feature through a Sigmoid activation layer to obtain a third extracted feature, and finally add the second extracted feature and the third extracted feature to obtain the physical guidance prediction result;

[0012] The prediction result refinement module includes five second encoders connected in sequence and five second decoders connected in sequence, and the second encoder and the second decoder of the same layer are correspondingly connected. The output feature of each layer of the second encoder and the output feature of the next layer of the second decoder are used as the input feature of the current layer of the second decoder. The output feature of the last layer of the second encoder and itself are used as the input feature of the current layer of the second decoder. The previous p-frame rainfall images, the advection prediction result and the physical guidance prediction result are spliced and used as the input feature of the first layer of the second encoder, and the output feature of the first layer of the second decoder is the final rainfall prediction result.

[0013] Preferably, the preprocessing is a normalization process, and the formula is as follows:

[0014] x = u×(4783 / 100)×12;

[0015] In the formula, u represents the radar echo rate of a single pixel point of the current frame image, and x represents the rainfall of a single pixel point of the current frame image.

[0016] Preferably, the advection simulator performs the following operations:

[0017] S2.1. Input the previous p-frame rainfall images of the rainfall image sequence into the first encoder to obtain a fourth extracted feature, and then input the fourth extracted feature into two parallel first decoders to respectively output the velocity field and the rainfall residual field of the subsequent q-frame rainfall images;

[0018] S2.2. Input the velocity fields of the p-th frame rainfall image and the (p + 1)-th frame rainfall image into an Euler integrator, and add the output result of the Euler integrator to the rainfall residual field of the (p + 1)-th frame rainfall image to obtain the (p + 1)-th frame rainfall image;

[0019] S2.3. Set p = p + 1, and return to execute step S2.2 until the subsequent q-frame rainfall images are obtained. Take the subsequent q-frame rainfall images as the advection prediction result.

[0020] Preferably, the Euler integrator adopts the forward and backward error compensation and correction method, and the formula is as follows:

[0021] X1 = Larangian(X, v, Δt);

[0022] X2 = Larangian(X1, v, -Δt);

[0023] 2e = X2 - X;

[0024]

[0025] In the formula, Larangian represents semi-Lagrangian integration, X represents the current-frame rainfall image, v represents the current velocity field, Δt represents the time interval between the current frame and the next frame, X1 represents the estimated rainfall image after Δt calculated by forward semi-Lagrangian integration, X2 represents the estimated rainfall image of the current frame calculated by backward semi-Lagrangian integration of X1, 2e represents the error between forward semi-Lagrangian integration and backward semi-Lagrangian integration, that is, the sum of the errors of the two semi-Lagrangian integrations. represents the estimated rainfall image after Δt after deducting the error, that is, the output result of the Euler integrator.

[0026] Preferably, during the pre-training process of the advection simulator or the near rainfall prediction network model, it is evaluated by the MSE loss, and the Adam algorithm is used to minimize the MSE loss to obtain the pre-trained advection simulator or near rainfall prediction network model. The MSE loss formula is as follows:

[0027]

[0028] In the formula, MSE represents the MSE loss, y t represents the true rainfall result of the t-th frame rainfall image, that is, the t-th frame rainfall image in the rainfall image sequence. represents the t-th frame rainfall image output by the advection simulator or the near rainfall prediction network model.

[0029] Preferably, the pre-trained advection simulator or the pre-trained near rainfall prediction network model is also compared and tested through two binary indexes and two non-binary indexes, where:

[0030] The binary indexes include the critical success index CSI and the Heidke skill score HSS, and the formulas are as follows:

[0031]

[0032] Wherein, TP is the true positive rate, TN is the true negative rate, FP is the false positive rate, and FN is the false negative rate;

[0033] The non-binarized metrics include the balanced mean squared error B-MSE and the balanced mean absolute error B-MAE, and the formulas are as follows:

[0034]

[0035] Wherein,

[0036]

[0037] In the formula, w t,i,j represents the balanced weight of the pixel point (i, j) of the rainfall image in the t-th frame, w(x) represents the balanced weight of a single pixel point in the rainfall image, and x t,i,j represents the rainfall amount of the pixel point (i, j) of the rainfall image in the t-th frame, represents the rainfall amount of the pixel point (i, j) of the rainfall image in the t-th frame in the final rainfall prediction result, H represents the height of the rainfall image in the final rainfall prediction result, and W represents the width of the rainfall image in the final rainfall prediction result.

[0038] Preferably, each encoder includes five first feature extraction modules connected in sequence. The first feature extraction module includes a max pooling layer, a convolutional layer, a normalization layer, a ReLU activation function, a convolutional layer, a normalization layer, and a ReLU activation function connected in sequence; each decoder includes five second feature extraction modules connected in sequence. The second feature extraction module includes a bilinear interpolation layer, a convolutional layer, a normalization layer, a ReLU activation function, a convolutional layer, a normalization layer, and a ReLU activation function connected in sequence.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0040] This method uses the first p frames of rainfall images in a rainfall image sequence to predict the next q frames of rainfall images based on a pre-trained advection simulator as the advection prediction result, and the advection prediction result includes the physical characteristics of the advection process. Then, the first p frames of rainfall images and the advection prediction result are input into a pre-trained near rainfall prediction network model based on physical constraints to obtain the final rainfall prediction result, realizing the prediction of rainfall. And a physical guidance prediction module is adopted in the near rainfall prediction network model based on physical constraints to constrain the strong rainfall data characteristics in the advection prediction result, and the final rainfall prediction result is obtained through refinement by the prediction result refinement module, so that the prediction error of the strong rainfall part is smaller. Compared with the prior art, it can generate a strong rainfall prediction result with smaller error, and physical constraints are introduced in this process, realizing the efficient prediction of near rainfall. Description of the Drawings

[0041] Figure 1 This is a flowchart of the near rainfall prediction method based on physical constraints of the present invention;

[0042] Figure 2 This is a structural block diagram of the advection simulator and the near rainfall prediction network model of the present invention;

[0043] Figure 3 This is a structural block diagram of the advection simulator of the present invention;

[0044] Figure 4 This is a schematic diagram of the true value, the velocity field predicted by the advection simulator, the rainfall residual field, and the advection prediction result of the present invention;

[0045] Figure 5 This is a schematic diagram for comparing the true value, the advection prediction result, and the NowcastNet evolution prediction result of the present invention;

[0046] Figure 6 This is a schematic diagram for comparing the final rainfall prediction results of the present invention and the existing model after five minutes. Detailed implementation manners

[0047] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0048] It should be noted that when a component is referred to as being "connected" to another component, it can be directly connected to the other component or there may also be an intermediate component. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field of the present application. The terms used in the description of the present application in this specification are only for the purpose of describing specific embodiments and are not intended to limit the present application.

[0049] As Figures 1-6 shown, a near rainfall prediction method based on physical constraints includes:

[0050] S1. Preprocess each frame of image in the meteorological radar echo sequence to obtain a corresponding rainfall image sequence.

[0051] In one embodiment, the preprocessing is normalization processing, and the formula is as follows:

[0052] x = u × (4783 / 100) × 12;

[0053] Among them, u represents the radar echo rate of a single pixel point in the current frame image, and x represents the rainfall of a single pixel point in the current frame image, and the unit of rainfall is mm / h.

[0054] S2. Predict the next q-frame rainfall images using the first p-frame rainfall images in the rainfall image sequence based on the pre-trained advection simulator as the advection prediction result, where p + q = N, and N is the total number of rainfall images in the rainfall image sequence.

[0055] In one embodiment, the advection simulator performs the following operations:

[0056] S2.1. Input the first p-frame rainfall images in the rainfall image sequence into the first encoder to obtain the fourth extracted feature, and then input the fourth extracted feature into two parallel first decoders to respectively output the velocity field and rainfall residual field of the next q-frame rainfall images;

[0057] S2.2. Input the velocity fields of the p-th frame rainfall image and the (p + 1)-th frame rainfall image into the Euler integrator, and add the output result of the Euler integrator to the rainfall residual field of the (p + 1)-th frame rainfall image to obtain the (p + 1)-th frame rainfall image;

[0058] S2.3. Set p = p + 1, and return to execute step S2.2 until the next q-frame rainfall images are obtained, and use the next q-frame rainfall images as the advection prediction result.

[0059] In this embodiment, each rainfall image sequence includes 18 frames, and p + q = 18. The values of p and q are variable. In this embodiment, p = 6 and q = 12 are taken as examples for description. Then the rainfall image sequence is expressed as {x1, x2,..., x 18}, and obtaining the first 6-frame data of the rainfall image sequence is expressed as {x1, x2,..., x6}, and the last 12-frame data of the rainfall image sequence is expressed as {x7, x8,..., x 18}.

[0060] 1) First, the first p-frame data of the rainfall image sequence pass through a first encoder and two first decoders (i.e., the velocity decoder and the residual decoder), and output the learnable velocity field and rainfall residual field of the next q frames.

[0061] The network model of the advection simulator constructed in the embodiment is as Figure 3 shown. The first encoder contains five layers connected in sequence. Each layer is composed of a max-pooling layer, a convolutional layer, a normalization layer, a ReLU activation function, a convolutional layer, a normalization layer, and a ReLU activation function in sequence. The output features enter the velocity decoder and the residual decoder respectively. The structures of the two first decoders are the same and contain five layers connected in sequence. Each layer is composed of a bilinear interpolation layer, a convolutional layer, a normalization layer, a ReLU activation function, a convolutional layer, a normalization layer, and a ReLU activation function in sequence.

[0062] The convolutional layer in the first encoder is a 2D convolution, and the convolutional layer (kernel size, stride, and padding width) is set to (3, 1, 1). The number of channels in the five 2D convolutional layers is sequentially set to 64, 128, 256, 512, and 512.

[0063] The size of the bilinear interpolation layer in the two first decoders is 2*2, and the convolutional layer is a 2D convolution. The convolutional layer (kernel size, stride, and padding width) is set to (3, 1, 1). The number of channels in the five 2D convolutional layers is sequentially set to 256, 128, 64, 64, and 12. Since it is desired to output a velocity field and a rainfall residual field of 12 frames, the number of channels in the last layer is set to 12.

[0064] In a specific embodiment, taking p = 6 and q = 12, the obtained 12-frame velocity field is represented as {v7, v8,..., v 18} and the 12-frame rainfall residual field is represented as {s7, s8,..., s 18}.

[0065] 2) Input the velocity fields of the p-th rainfall image and the (p + 1)-th rainfall image into the Euler integrator, and add the output result of the Euler integrator to the rainfall residual field of the (p + 1)-th rainfall image to obtain the (p + 1)-th rainfall image. The subsequent q-frame rainfall images are obtained through frame-by-frame integration as the advection prediction result.

[0066] The Euler integrator mainly uses the forward and backward error compensation and correction method for integration. The input is the rainfall of the current frame (the current-frame rainfall image) and the velocity field of the next frame, and the rainfall of the next frame (the next-frame rainfall image) is obtained. The forward and backward error compensation and correction method mainly first performs forward simulation, using the current-frame rainfall image x t , the velocity field v of the next frame t+1 , to predict the estimated rainfall image after a time interval of Δt Then, the estimated rainfall image is subjected to backward simulation to obtain the estimated rainfall image Δt before Assuming no prediction error, i.e., the error e = 0, it can be deduced that However, in reality, errors will be obtained after forward simulation and backward simulation In subsequent integration, after removing the error e, a more accurate rainfall prediction result can be obtained Specifically as follows:

[0067] In one embodiment, the Euler integrator adopts the forward and backward error compensation and correction method, and the formula is as follows:

[0068] X1 = Larangian(X, v, Δt);

[0069] X2 = Larangian(X1, v, -Δt);

[0070] 2e = X2 - X;

[0071]

[0072] Wherein, Larangian represents the semi-Lagrangian integral, X represents the current frame rainfall image, v represents the current velocity field, Δt represents the time interval between the current frame and the next frame, X1 represents the estimated rainfall image after Δt obtained by forward semi-Lagrangian integration, X2 represents the estimated rainfall image of the current frame obtained by X1 through reverse semi-Lagrangian integration, and 2e represents the error between the forward semi-Lagrangian integration and the reverse semi-Lagrangian integration, that is, the sum of the errors of the two semi-Lagrangian integrations. represents the estimated rainfall image after Δt after deducting the error, that is, the output result of the Euler integrator.

[0073] In a specific embodiment, p = 6 and q = 12 are taken. First, the input of the Euler integrator is the 6th frame rainfall image x6 and the 7th frame velocity field v7. The output result of the Euler integrator and the 7th frame rainfall residual field s7 are used to obtain the 7th frame rainfall image. Then, the corresponding rainfall image, velocity field, and rainfall residual field are input frame by frame to obtain the rainfall image of the next frame, and the iteration is repeated until the advection prediction result of the last 12 frames of data in the rainfall image sequence is obtained. The Euler integrator can be the same one.

[0074] It should be noted that in the assumption of the advection process, since the divergence of the rainfall image is 0, that is, the overall rainfall intensity does not change in magnitude, only moves, so the velocity field {v7, v8,..., v 18} is estimated based on the principle of the advection process, and there is no change in the overall rainfall amount obtained through the Euler integrator in terms of rainfall intensity magnitude. To estimate the change in rainfall intensity magnitude, the rainfall residual field {s7, s8,..., s 18} is used as the value of the rainfall intensity change at each pixel point, and is superimposed on the result after integrating using the velocity field to obtain the final advection prediction result.

[0075] S3. Take the advection prediction result as a physical constraint and combine it with the first p frame rainfall images, and input it into the pre-trained adjacent rainfall prediction network model to obtain the final rainfall prediction result. The adjacent rainfall prediction network model includes a physical guidance prediction module and a prediction result refinement module, where:

[0076] The physical guidance prediction module includes an advection prediction result feature extraction module and a gate mechanism module. The advection prediction result feature extraction module is used to separately extract the maximum value and the average value of the advection prediction result through a parallel first calculation module and second calculation module, and add the maximum value and the average value to obtain a first extraction feature. The first calculation module includes a global max pooling layer and a fully connected layer connected in sequence, and the second calculation module includes a global average pooling layer and a fully connected layer connected in sequence;

[0077] The gate mechanism module is used to splice the previous p-frame rainfall images and the advection prediction result to obtain a spliced feature, convert the spliced feature through a convolutional layer into a second extraction feature with the same dimension as the advection prediction result, multiply the second extraction feature by the first extraction feature through a Sigmoid activation layer to obtain a third extraction feature, and finally add the second extraction feature and the third extraction feature to obtain the physical guidance prediction result;

[0078] The prediction result refinement module includes five second encoders connected in sequence and five second decoders connected in sequence, and the second encoder and the second decoder of the same layer are correspondingly connected. The output feature of each layer of the second encoder and the output feature of the next layer of the second decoder are used as the input feature of the current layer of the second decoder. The output feature of the last layer of the second encoder and itself are used as the input feature of the current layer of the second decoder. The previous p-frame rainfall images, the advection prediction result and the physical guidance prediction result are spliced and used as the input feature of the first layer of the second encoder, and the output feature of the first layer of the second decoder is the final rainfall prediction result.

[0079] Among them, as Figure 2 shown, the advection prediction result feature extraction module consists of two calculation modules. One calculation module consists of a global max pooling layer (global max pooling) and a fully connected layer connected in sequence, and the other calculation module consists of a global average pooling layer (global average pooling) and a fully connected layer connected in sequence. The two calculation modules respectively obtain the maximum value and the average value of the advection prediction result, and sum the two to obtain the data feature (first extraction feature) of the advection prediction result. The number of input features of both fully connected layers is 12, and the number of output features is 12.

[0080] The gate mechanism module consists of a convolutional layer and a Sigmoid activation layer (Sigmoid function) connected in sequence. The total dimension 18 is converted into the prediction result dimension 12 through the convolutional layer to obtain the overall feature (second extraction feature) of the advection prediction result. The Sigmoid activation layer is a kind of gate mechanism. The rainfall data is multiplied by the data feature of the advection prediction result through the Sigmoid activation layer to determine the finally retained feature (third extraction feature). The overall feature of the advection prediction result and the third extraction feature are added to obtain the prediction result of the physical guidance prediction module.

[0081] The prediction result refinement module consists of five sequentially connected second encoders and five sequentially connected second decoders. Each layer in the second encoder consists of a max pooling layer, a convolutional layer, a normalization layer, a ReLU activation function, a convolutional layer, a normalization layer, and a ReLU activation function in sequence. The convolutional layer is a 2D convolution, and the convolutional layer (kernel size, convolution stride, and padding width) is set to (3, 1, 1). The number of channels of the 2D convolutional layers in the five second encoders is set to 64, 128, 256, 512, and 512 in sequence. Each layer in the second decoder is a bilinear interpolation layer, a convolutional layer, a normalization layer, a ReLU activation function, a convolutional layer, a normalization layer, and a ReLU activation function in sequence. The size of the bilinear interpolation layer in the second decoder is 2*2. The convolutional layer is a 2D convolution, and the 2D convolutional layer (kernel size, convolution stride, and padding width) is set to (3, 1, 1). The number of channels of the 2D convolutional layers in the second decoder is set to 256, 128, 64, 64, and 12 in sequence. Since it is desired to output 12 frames of rainfall images, the number of channels of the last layer is set to 12. At the same time, there is a direct connection between the same layers of the second encoder and the second decoder. The output features of each layer of the second encoder and the output features of the next layer of the second decoder are used as the input features of the current layer of the second decoder. The output features of the last layer of the second encoder and itself are used as the input features of the current layer of the second decoder.

[0082] By inputting the rainfall images of the first p frames (the first 6 frames) and the advection prediction results of the next q frames (the next 12 frames) into the constructed and pre-trained short-term rainfall prediction network model that focuses on heavy rainfall, the final rainfall prediction result is obtained. Specifically:

[0083] The advection prediction results of the next 12 frames are respectively input into the advection prediction result feature extraction module, and then the results of the two calculation modules are added to obtain the first calculation result (the first extracted feature, that is, the feature data of the advection prediction result with 12 channels);

[0084] The rainfall images of the first 6 frames and the advection prediction results of the next 12 frames are input into the gate mechanism module. In the gate mechanism module, the rainfall images of the first 6 frames are concatenated with the advection prediction results of the next 12 frames (a total of 18 frames of data). The output result of the convolutional layer is the second calculation result (the dimension is converted from 18 to 12, that is, the second extracted feature), and the output result of the Sigmoid activation layer is the third calculation result (the dimension is 12);

[0085] Specifically, it is hoped that the third calculation result in the gate mechanism module can represent the overall features of the first 6 frames of data and the advection prediction results of the next 12 frames to constrain the data features of the advection prediction results.

[0086] Multiply the first calculation result by the third calculation result to obtain a fourth calculation result (feature data passing through the gate mechanism with a dimension of 12), and add the fourth calculation result to the second calculation result to obtain a physical guidance prediction result (with a dimension of 12);

[0087] Specifically, the product of the first calculation result and the third calculation result is the feature filtered by the gate mechanism, and the feature data that needs more attention is superimposed on the first calculation result. To a certain extent, the first calculation result represents the overall features of the first 6 frames of data and the advection prediction results of the last 12 frames.

[0088] Finally, splice the rainfall data of the first 6 frames of rainfall images, the physical guidance prediction result (12 frames), and the advection prediction result (12 frames) (30 frames of data) and input them into the first-layer second encoder of the prediction result refinement module for feature data extraction. Through the prediction result refinement module, obtain the refined rainfall prediction result of the last 12 frames, which is the final rainfall prediction result.

[0089] In one embodiment, during the pre-training process of the advection simulator or the adjacent rainfall prediction network model, it is evaluated through the MSE loss, and the Adam algorithm is used to minimize the MSE loss to obtain a pre-trained advection simulator or adjacent rainfall prediction network model. The MSE loss formula is as follows:

[0090]

[0091] In the formula, MSE represents the MSE loss, and y t represents the true rainfall result of the t-th frame of the rainfall image, that is, the t-th frame of the rainfall image in the rainfall image sequence, represents the t-th frame of the rainfall image output by the advection simulator or the adjacent rainfall prediction network model.

[0092] During the training process of the advection simulator or the adjacent rainfall prediction network model, the dataset used is several radar echo maps, each image with a pixel size of 288×288, a time resolution of 5 minutes for the image, that is, the interval between each frame of the image is 5 minutes, and the unit of the pixel value in each image is mm / 5 minutes. The dataset randomly divides all the radar echo maps, resulting in 4000 sequences in the training set, 1557 sequences in the test set, and 1734 sequences in the validation set. For example, the training of the adjacent rainfall prediction network model is carried out on the training set. The initial weight values of this network model are randomly selected, with a learning rate of 0.0001. In this embodiment, a total of 100 complete trainings are carried out. And each time after a training, it is verified on the validation set to obtain a loss value. The loss value is calculated by calculating the MSE loss between the final rainfall prediction result and the last q frames of data (true rainfall data, that is, the true value) in the rainfall image sequence of the validation set, and the Adam algorithm is used to minimize the loss. Finally, a pre-trained adjacent rainfall prediction network model is obtained.

[0093] In one embodiment, the pre-trained advection simulator or the pre-trained near rainfall prediction network model is also compared and tested through two binary metrics and two non-binary metrics, where:

[0094] The binary metrics include the Critical Success Index (CSI) and the Heidke Skill Score (HSS), and the formulas are as follows:

[0095]

[0096] In the formula, TP is the true positive rate, TN is the true negative rate, FP is the false positive rate, and FN is the false negative rate;

[0097] The non-binary metrics include the Balanced Mean Squared Error (B-MSE) and the Balanced Mean Absolute Error (B-MAE), and the formulas are as follows:

[0098]

[0099] Where,

[0100]

[0101] In the formula, w t,i,j represents the balanced weight of the pixel point (i, j) of the t-th frame rainfall image, w(x) represents the balanced weight of a single pixel point in the rainfall image, and x t,i,j represents the rainfall amount of the pixel point (i, j) of the t-th frame rainfall image, represents the rainfall amount of the pixel point (i, j) of the t-th frame rainfall image in the final rainfall prediction result, H represents the height of the rainfall image in the final rainfall prediction result, and W represents the width of the rainfall image in the final rainfall prediction result.

[0102] Among them, the pre-trained model is tested on the test set, and the comparison between the model prediction result and the corresponding real rainfall data is mainly through 2 binary metrics and 2 non-binary metrics. The binary metrics will compare the two according to the set rainfall intensity threshold, and the non-binary metrics are used to evaluate the similarity between the two from another perspective.

[0103] The binarization metrics used are the Critical Success Index (CSI) and the Heidke Skill Score (HSS). CSI measures the ratio of the number of correctly predicted samples to the sum of all correctly and incorrectly predicted samples, and HSS measures the improvement of the prediction compared to a random prediction. The binarization metrics are mainly calculated through the following four metrics: True Positives (TP), True Negatives (TN), False Positives (FP), and False Negatives (FN).

[0104] In this embodiment, the rainfall intensity is divided into no rain, light rain, light rain to moderate rain, moderate rain, heavy rain and above, corresponding to rainfall amounts of 0.5 mm / h, 2 mm / h, 5 mm / h, and 10 mm / h as thresholds respectively, to binarize the prediction results and the actual rainfall conditions. For the CSI and HSS metrics, the larger the metric value, the more similar the prediction result is to the true value.

[0105] Due to the problem of extremely unbalanced data in the dataset, specifically, the data of no rain in the dataset is much more than the dataset with rainfall, and the number of heavy rain occurrences is also less than that of light rain occurrences. Therefore, balance weights are introduced in the non-binarization metrics Balance-Mean Squared Error (B-MSE) and Balance-Mean Absolute Error (B-MAE), and different weights are assigned according to the rainfall intensity. For the non-binarization metrics B-MSE and B-MAE, the lower the metric value, the smaller the error between the prediction result and the actual rainfall condition, and the better the prediction effect, that is, the closer the prediction result is to the true value.

[0106] The advection simulator part is trained separately, using B-MAE as the loss function, with random values for the initial weights, and a learning rate of 0.0001. In this embodiment, a total of 100 complete trainings are carried out. The obtained prediction results are shown in Figure 4 It can be seen that the advection simulator has learned the overall direction of rainfall movement and the local intensity change. And the advection prediction results of this part are compared with the evolution prediction results of the evolution prediction method in NowcastNet (NowcastNet evolution prediction results). As shown in Figure 5 The advection prediction results can predict rainfall better than the evolution prediction results, and after the prediction time increases, the predicted rainfall area will not spread and the predicted image will not be blurred. Among them, T refers to starting from the predicted starting moment, that is, the time point of the 6th frame of rainfall image.

[0107] Compare the adjacent rainfall prediction network model (PiCNet) in this method with the existing relatively good models, namely ConvLSTM (Convolutional Long Short-Term Memory Network), PredRNN (Recurrent Neural Network for Spatiotemporal Predictive Learning), PredRNN++ (the enhanced version of PredRNN), PFST, and MIM (Memory in memory, a neural network for learning high-order non-linear spatiotemporal dynamics), in terms of four evaluation metrics, as shown in Table 1:

[0108] Table 1

[0109]

[0110]

[0111] As can be seen from Table 1, PiCNet is superior to the existing models in terms of the evaluation metrics HSS, CSI, and B-MSE for rainfall intensities of light rain and above, and B-MAE is basically on par with the existing models. Especially in the heavy rainfall part (x ≥ 10), HSS and CSI have a significant improvement compared to other models, and the overall prediction accuracy has a significant improvement.

[0112] The prediction results of PiCNet, ConvLSTM, PredRNN, PredRNN++, and PFST for the rainfall situation in the next five minutes are shown in Figure 6 . Comparing with the true value Groundtruth of the rainfall situation in the next five minutes, it can be seen that the prediction results of PiCNet are better than those of other models. Especially in the heavy rainfall area, more accurate prediction results can be obtained.

[0113] Among them, the evolution prediction method in NowcastNet refers to the technical solution in Zhang Y, Long M, Chen K, et al. Skilful nowcasting of extreme precipitation with NowcastNet [J]. Nature, 2023, 619(7970): 526-532; ConvLSTM refers to the deep learning method disclosed in Shi X, Chen Z, Wang H, et al. Convolutional LSTM network: A machine learning approach for precipitation nowcasting [J]. Advances in neural information processing systems, 2015, 28; PredRNN refers to the deep learning method disclosed in Wang Y, Long M, Wang J, et al. Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms [J]. Advances in neural information processing systems, 2017, 30; PredRNN++ refers to the deep learning method disclosed in Wang Y, Gao Z, Long M, et al. Predrnn++: Towards a resolution of the deep-in-time dilemma in spatiotemporal predictive learning [C] / / International conference on machine learning. PMLR, 2018: 5123-5132; PFST refers to the deep learning method disclosed in Luo C, Li X, Ye Y. PFST-LSTM: A spatiotemporal LSTM model with pseudoflow prediction for precipitation nowcasting [J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2020, 14: 843-857.

[0114] The technical features of the above-described embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as falling within the scope described in this specification.

[0115] The above-described embodiments only represent relatively specific and detailed embodiments of the present application, but should not be construed as limiting the scope of the patent application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. A method for predicting approaching rainfall based on physical constraints, characterized in that: The near rainfall prediction method based on physical constraints includes the following steps: S1. Preprocess each frame of the meteorological radar echo sequence to obtain a corresponding rainfall image sequence; S2. Based on the pre-trained advection simulator, use the first p frames of rainfall images in the rainfall image sequence to predict the next q frames of rainfall images as the advection prediction result, where p + q = N, and N is the total number of rainfall images in the rainfall image sequence; The advection simulator performs the following operations: S2.

1. Input the first p frames of rainfall images in the rainfall image sequence into the first encoder to obtain the fourth extraction feature, and then input the fourth extraction feature into two parallel first decoders to respectively output the velocity field and rainfall residual field of the next q frames of rainfall images; S2.

2. Input the velocity fields of the p-th frame of rainfall image and the (p + 1)-th frame of rainfall image into the Euler integrator, and add the output result of the Euler integrator to the rainfall residual field of the (p + 1)-th frame of rainfall image to obtain the (p + 1)-th frame of rainfall image; S2.

3. Set p = p + 1, return to execute step S2.2 until the next q frames of rainfall images are obtained, and use the next q frames of rainfall images as the advection prediction result; S3. Use the advection prediction result as a physical constraint and combine it with the first p frames of rainfall images, and input them into the pre-trained near rainfall prediction network model to obtain the final rainfall prediction result. The near rainfall prediction network model includes a physical guidance prediction module and a prediction result refinement module, where: The physical guidance prediction module includes an advection prediction result feature extraction module and a gate mechanism module. The advection prediction result feature extraction module is used to respectively extract the maximum value and average value of the advection prediction result through a parallel first calculation module and second calculation module, and add the maximum value and average value to obtain the first extraction feature. The first calculation module includes a global max pooling layer and a fully connected layer connected in sequence. The second calculation module includes a global average pooling layer and a fully connected layer connected in sequence; The gate mechanism module is used to splice the first p frames of rainfall images and the advection prediction result to obtain a spliced feature, convert the spliced feature through a convolutional layer into a second extraction feature with the same dimension as the advection prediction result, multiply the second extraction feature by the first extraction feature through a Sigmoid activation layer to obtain a third extraction feature, and finally add the second extraction feature and the third extraction feature to obtain the physical guidance prediction result; The prediction result refinement module includes five second encoders connected in sequence and five second decoders connected in sequence, and the second encoder and second decoder of the same layer are correspondingly connected. The output feature of each layer of the second encoder and the output feature of the next layer of the second decoder are used as the input feature of the current layer of the second decoder. The output feature of the last layer of the second encoder and itself are used as the input feature of the current layer of the second decoder. Splice the first p frames of rainfall images, the advection prediction result and the physical guidance prediction result and use them as the input feature of the first layer of the second encoder. The output feature of the first layer of the second decoder is the final rainfall prediction result.

2. The method for predicting approaching rainfall based on physical constraints according to claim 1, characterized in that: The preprocessing is normalization processing, and the formula is as follows: x = u × (4783 / 100) × 12; Wherein, u represents the radar echo rate of a single pixel point of the current frame image, and x represents the rainfall of a single pixel point of the current frame image.

3. The method for predicting approaching rainfall based on physical constraints according to claim 1, wherein: The Euler integrator adopts a forward and backward error compensation and correction method, and the formula is as follows: X1 = Larangian(X, v, Δt); X2 = Larangian(X1, v, -Δt); 2e = X2 - X; In the formula, the Lagrangian represents the semi-Lagrangian integral, X represents the current frame rainfall image, v represents the current velocity field, Δt represents the time interval between the current frame and the next frame, X1 represents the estimated rainfall image after Δt calculated by the forward semi-Lagrangian integral, X2 represents the estimated rainfall image of the current frame calculated by X1 through the backward semi-Lagrangian integral, 2e represents the error between the forward semi-Lagrangian integral and the backward semi-Lagrangian integral, that is, the sum of the errors of the two semi-Lagrangian integrals. represents the estimated rainfall image after Δt after deducting the error, that is, the output result of the Euler integrator.

4. The method for predicting approaching rainfall based on physical constraints according to claim 1, wherein: During the pre-training process of the advection simulator or the nowcasting rainfall prediction network model, it is evaluated through the MSE loss, and the Adam algorithm is used to minimize the MSE loss to obtain the pre-trained advection simulator or the nowcasting rainfall prediction network model. The MSE loss formula is as follows: where MSE represents the MSE loss, and y t represents the true rainfall result of the t-th frame rainfall image, that is, the t-th frame rainfall image in the rainfall image sequence, represents the t-th frame rainfall image output by the advection simulator or the adjacent rainfall prediction network model.

5. The method for predicting approaching rainfall based on physical constraints according to claim 4, characterized in that: The pre-trained advection simulator or the pre-trained nowcasting rainfall prediction network model is also compared and tested through two binary indexes and two non-binary indexes, where: The binary indexes include the critical success index CSI and the Heidke skill score HSS, and the formulas are as follows: Wherein, TP is the true positive rate, TN is the true negative rate, FP is the false positive rate, and FN is the false negative rate; The non-binary indexes include the balanced mean square error B-MSE and the balanced mean absolute error B-MAE, and the formulas are as follows: Wherein, where w t,i,j represents the balance weight of the pixel point (i, j) in the rainfall image of the t-th frame, w(x) represents the balance weight of a single pixel point in the rainfall image, and x t,i,j represents the rainfall of the pixel point (i, j) in the rainfall image of the t-th frame, represents the rainfall of the pixel point (i, j) in the rainfall image of the t-th frame in the final rainfall prediction result, H represents the height of the rainfall image in the final rainfall prediction result, and W represents the width of the rainfall image in the final rainfall prediction result.

6. The method for predicting approaching rainfall based on physical constraints according to any one of claims 1 to 5, characterized in that: Each of the encoders includes five first feature extraction modules connected in sequence. The first feature extraction module includes a max pooling layer, a convolutional layer, a normalization layer, a ReLU activation function, a convolutional layer, a normalization layer, and a ReLU activation function connected in sequence; each of the decoders includes five second feature extraction modules connected in sequence. The second feature extraction module includes a bilinear interpolation layer, a convolutional layer, a normalization layer, a ReLU activation function, a convolutional layer, a normalization layer, and a ReLU activation function connected in sequence.

Images