Adjacent rainfall prediction method based on physical constraint

Abstract

The invention discloses an adjacent rainfall prediction method based on physical constraints. The method comprises the following steps: S1, preprocessing each frame of image in a meteorological radar echo sequence to obtain a corresponding rainfall image sequence; s2, on the basis of a pre-trained advection simulator, the last p frames of rainfall images of the rainfall image sequence are adopted to predict the last q frames of rainfall images to serve as advection prediction results, p + q = N, and N is the total frame number of the rainfall images in the rainfall image sequence; s3, the advection prediction result serves as physical constraint, the advection prediction result is combined with the first p frames of rainfall images and input into a pre-trained proximity rainfall prediction network model to obtain a final rainfall prediction result, and the proximity rainfall prediction network model comprises a physical guidance prediction module and a prediction result refining module. According to the method, a heavy rainfall prediction result with a smaller error can be generated, physical constraints are introduced in the process, and high-efficiency prediction of approaching rainfall is realized.

Description

Technical Field

[0001] The invention belongs to the field of rainfall prediction, and in particular relates to a near-term rainfall prediction method based on physical constraints. Background Art

[0002] Recently, meteorological disasters have occurred frequently in various places. For example, extreme precipitation, a common meteorological phenomenon, can cause meteorological disasters such as floods and mudslides, resulting in significant social, ecological and economic losses. Therefore, rainfall forecasting research has very important practical significance.

[0003] At present, meteorological agencies that perform rainfall forecasts mainly use numerical calculation methods. However, due to the need to use a large amount of diverse meteorological data to simulate the initial field for forecasting, traditional numerical calculation methods have the problems of difficulty in data acquisition and high consumption of computing resources. Deep learning methods have gradually developed in this context. The rainfall forecasting method based on deep learning is a data-driven method. Compared with traditional methods, it requires smaller data and fewer computing resources. However, most models are black box models that rely solely on data-driven and ignore the importance of the 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 near-intensity rainfall forecasting method based on physical constraints. Summary of the invention

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

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] The present invention proposes a method for imminent rainfall prediction based on physical constraints, comprising the following steps:

[0007] S1, preprocessing each frame image in the meteorological radar echo sequence to obtain a corresponding rainfall image sequence;

[0008] S2, based on the pre-trained advection simulator, the first p frames of rainfall images in the rainfall image sequence are used to predict the last q frames of rainfall images as the advection prediction result, where p+q=N, N is the total number of rainfall image frames in the rainfall image sequence;

[0009] S3. The advection prediction result is combined with the previous p frames of rainfall images as a physical constraint and input into the pre-trained near-term rainfall prediction network model to obtain the final rainfall prediction result. The near-term rainfall prediction network model includes a physical guidance prediction module and a prediction result refinement module, wherein:

[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 extract the maximum value and the average value of the advection prediction result respectively through a first calculation module and a second calculation module in parallel, and add the maximum value and the average value to obtain a first extracted feature. The first calculation module includes a global maximum 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 rainfall image of the previous p frames and the advection prediction results to obtain the splicing features, and convert the splicing features into the second extracted features of the same dimension as the advection prediction results through the convolution layer, and then multiply the second extracted features with the first extracted features through the Sigmoid activation layer to obtain the third extracted features, and finally add the second extracted features to the third extracted features to obtain the physical guidance prediction results;

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

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

[0014] x = u × (4783 / 100) × 12;

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

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

[0017] S2.1, inputting the first p frames of rainfall images of the rainfall image sequence into the first encoder to obtain the fourth extracted features, and then inputting the fourth extracted features into two parallel first decoders to output the velocity field and rainfall residual field of the next q frames of rainfall images respectively;

[0018] S2.2, inputting the velocity fields of the p-th frame rainfall image and the p+1-th frame rainfall image into the Euler integrator, and adding 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 next q frames of rainfall images are obtained, and the next q frames of rainfall images are used as advection prediction results.

[0020] Preferably, the Euler integrator adopts a 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] Where Langrangian represents the semi-Lagrangian integral, X represents the rainfall image of the current frame, 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 integral calculation, X2 represents the estimated rainfall image of the current frame obtained by reverse semi-Lagrangian integral calculation of X1, 2e represents the error between the forward semi-Lagrangian integral and the reverse semi-Lagrangian integral, that is, the error sum of the two semi-Lagrangian integrals, It represents the estimated rainfall image after deducting the error Δt, that is, the output result of the Euler integrator.

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

[0027]

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

[0029] Preferably, the pre-trained advection simulator or the pre-trained near-falling rainfall prediction network model is also subjected to comparative testing by two binary indices and two non-binary indices, wherein:

[0030] The binary indicators include the critical success index CSI and the Heidecke skill score HSS, and the formula is as follows:

[0031]

[0032] 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;

[0033] The non-binarization indicators include balanced mean square error B-MSE and balanced mean absolute error B-MAE, the formulas are as follows:

[0034]

[0035] in,

[0036]

[0037] In the formula, w t,i,j represents the balance weight of the pixel (i, j) in the t-th frame rainfall image, w(x) represents the balance weight of a single pixel in the rainfall image, x t,i,j represents the rainfall at pixel (i, j) of the t-th frame rainfall image, represents the rainfall at 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.

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

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] The method adopts the first p-frame rainfall images of a rainfall image sequence to predict the last q-frame rainfall images as the advection prediction result based on a pre-trained advection simulator, and the advection prediction result includes the physical characteristics of the advection process, and then the first p-frame rainfall images and the advection prediction result are input into a pre-trained near-falling rainfall prediction network model based on physical constraints to obtain the final rainfall prediction result, so as to realize the prediction of rainfall, and adopts a physical guidance prediction module in the near-falling rainfall prediction network model based on physical constraints to constrain the heavy rainfall data features in the advection prediction result, and refines the final rainfall prediction result through the prediction result refinement module, so that the prediction error of the heavy rainfall part is smaller, and compared with the prior art, it can produce a heavy rainfall prediction result with smaller error, and in this process, physical constraints are introduced to realize the efficient prediction of near-falling rainfall. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a flow chart of the imminent rainfall prediction method based on physical constraints of the present invention;

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

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

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

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

[0046] Figure 6 The figure is a schematic diagram comparing the final rainfall prediction results after five minutes between the present invention and the existing model. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

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

[0049] like Figure 1-6 As shown, a method for nowcasting rainfall based on physical constraints includes:

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

[0051] In one embodiment, the preprocessing is a standardization process, and the formula is as follows:

[0052] x = u × (4783 / 100) × 12;

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

[0054] S2. Based on the pre-trained advection simulator, the first p frames of rainfall images in the rainfall image sequence are used to predict the last q frames of rainfall images as the advection prediction result, wherein p+q=N, and N is the total number of rainfall image frames in the rainfall image sequence.

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

[0056] S2.1, inputting the first p frames of rainfall images of the rainfall image sequence into the first encoder to obtain the fourth extracted features, and then inputting the fourth extracted features into two parallel first decoders to output the velocity field and rainfall residual field of the next q frames of rainfall images respectively;

[0057] S2.2, inputting the velocity fields of the p-th frame rainfall image and the p+1-th frame rainfall image into the Euler integrator, and adding 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 frames of rainfall images are obtained, and the next q frames of rainfall images are used as advection prediction results.

[0059] In this embodiment, each rainfall image sequence includes 18 frames, and p+q=18, the values ​​of p and q are variable. This embodiment is described by taking p=6 and q=12 as an example, and the rainfall image sequence is represented as {x1, x2, …, x 18}, the first 6 frames of the rainfall image sequence are expressed as {x1, x2, …, x6}, and the last 12 frames of the rainfall image sequence are expressed as {x7, x8, …, x 18}.

[0060] 1) First, the first p frames of the rainfall image sequence are passed through a first encoder and two first decoders (i.e., a velocity decoder and a residual decoder) to 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 follows Figure 3 As shown in the figure, the first encoder includes five layers connected in sequence, each of which is composed of a maximum pooling layer, a convolutional layer, a normalization layer, a ReLU activation function, a convolutional layer, a normalization layer, and a ReLU activation function. The output features enter the speed decoder and the residual decoder respectively. The two first decoders have the same structure, including five layers connected in sequence, each of which 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.

[0062] The convolutional layers in the first encoder are 2D convolutions, and the convolutional layers (convolution kernel size, convolution step, and padding width) are all set to (3, 1, 1). The number of channels of the five 2D convolutional layers is set to 64, 128, 256, 512, and 512 respectively.

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

[0064] In a specific embodiment, p=6, q=12, and 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) The velocity fields of the p-th frame rainfall image and the p+1-th frame rainfall image are input into the Euler integrator, and the output result of the Euler integrator is added to the rainfall residual field of the p+1-th frame rainfall image to obtain the p+1-th frame rainfall image. The next q frames of rainfall images are obtained by 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 performs forward simulation first, using the current frame rainfall image x t , the next frame velocity field v t+1 , predict the estimated rainfall image after time Δt Then the estimated rainfall image Perform reverse simulation to obtain the estimated rainfall image before time Δt Assuming that there is no prediction error, that is, error e = 0, it can be deduced that But in fact, after forward simulation and reverse simulation, errors will be obtained After removing the error e in the subsequent integration, a more accurate rainfall prediction result can be obtained. The details are as follows:

[0067] In one embodiment, the Euler integrator uses a 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] Where Langrangian represents the semi-Lagrangian integral, X represents the rainfall image of the current frame, 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 integral calculation, X2 represents the estimated rainfall image of the current frame obtained by reverse semi-Lagrangian integral calculation of X1, 2e represents the error between the forward semi-Lagrangian integral and the reverse semi-Lagrangian integral, that is, the error sum of the two semi-Lagrangian integrals, It represents the estimated rainfall image after deducting the error Δt, that is, the output result of the Euler integrator.

[0073] In a specific embodiment, p=6, q=12, first, the input of the Euler integrator is the 6th frame rainfall image x6, the 7th frame velocity field v7, and the output result of the Euler integrator and the 7th frame rainfall residual field s7 are combined 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 results of the last 12 frames of rainfall image sequence are obtained. The Euler integrator can be the same.

[0074] It should be noted that in the advection process, since the divergence of the rainfall image is 0, that is, the overall rainfall intensity does not change, only moves, so the velocity field {v7,v8,…,v 18 Based on the principle of advection process, the rainfall obtained by Euler integrator does not have any change in rainfall intensity. In order to estimate the change in rainfall intensity, the rainfall residual field {s7,s8,…,s 18} is used as the value of rainfall intensity change at each pixel, and superimposed on the result after integration using the velocity field to obtain the final advection prediction result.

[0075] S3. The advection prediction result is combined with the previous p frames of rainfall images as a physical constraint and input into the pre-trained near-term rainfall prediction network model to obtain the final rainfall prediction result. The near-term rainfall prediction network model includes a physical guidance prediction module and a prediction result refinement module, wherein:

[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 extract the maximum value and the average value of the advection prediction result respectively through a first calculation module and a second calculation module in parallel, and add the maximum value and the average value to obtain a first extracted feature. The first calculation module includes a global maximum 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 rainfall image of the previous p frames and the advection prediction results to obtain the splicing features, and convert the splicing features into the second extracted features of the same dimension as the advection prediction results through the convolution layer, and then multiply the second extracted features with the first extracted features through the Sigmoid activation layer to obtain the third extracted features, and finally add the second extracted features to the third extracted features to obtain the physical guidance prediction results;

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

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

[0080] The gate mechanism module is composed of a convolution layer and a Sigmoid activation layer (Sigmoid function) in sequence. The convolution layer converts the total dimension 18 into the prediction result dimension 12 to obtain the overall characteristics of the advection prediction result (the second extracted feature). The Sigmoid activation layer is a gate mechanism. The rainfall data is multiplied by the data characteristics of the advection prediction result through the Sigmoid activation layer to determine the final retained features (the third extracted features). The overall characteristics of the advection prediction result and the third extracted features are added to obtain the prediction results of the physical guidance prediction module.

[0081] The prediction result refinement module consists of five layers of second encoders connected in sequence and five layers of second decoders connected in sequence. Each layer in the second encoder is composed of a maximum pooling layer, a convolution layer, a normalization layer, a ReLU activation function, a convolution layer, a normalization layer, and a ReLU activation function. The convolution layer is a 2D convolution. The convolution layer (convolution kernel size, convolution step size, and padding width) are all set to (3, 1, 1). The number of channels of the 2D convolution layer in the five-layer second encoder is set to 64, 128, 256, 512, and 512, respectively; each layer in the second decoder is a bilinear interpolation layer, a convolution layer, a normalization layer, a ReLU activation function, a convolution layer, a normalization layer, and a ReLU activation function. The size of the bilinear interpolation layer in the second decoder is 2*2, the convolution layer is a 2D convolution, and the 2D convolution layer (convolution kernel size, convolution step size and padding width) are all set to (3, 1, 1). The number of channels of the 2D convolution layer in the second decoder is set to 256, 128, 64, 64 and 12 respectively. Since the output is expected to be 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 second encoder and the second decoder in the same layer. The output features of the second encoder of each layer and the output features of the second decoder of the next layer are used as the input features of the second decoder of the current layer, and the output features of the second encoder of the last layer and itself are used as the input features of the second decoder of the current layer.

[0082] The final rainfall prediction result is obtained by inputting the advection prediction results of the first p frames (first 6 frames) rainfall images and the last q frames (last 12 frames) into the constructed and pre-trained near-fall rainfall prediction network model focusing on heavy rainfall. Specifically:

[0083] The advection prediction results of the last 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, i.e., the feature data of the advection prediction result with the number of channels being 12);

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

[0085] Specifically, the process hopes that the third calculation result in the gate mechanism module can represent the overall characteristics of the first 6 frames of data and the last 12 frames of advection prediction results to constrain the data characteristics of the advection prediction results.

[0086] The first calculation result is multiplied by the third calculation result to obtain a fourth calculation result (characteristic data of the gate mechanism with a dimension of 12), and the fourth calculation result is added to the second calculation result to obtain a physical guidance prediction result (with a dimension of 12);

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

[0088] Finally, the rainfall data of the first 6 frames of rainfall images, the physical guidance prediction results (12 frames) and the advection prediction results (12 frames) are spliced ​​(30 frames of data) and input into the first-layer second encoder of the prediction result refinement module for feature data extraction. The refined rainfall prediction results of the last 12 frames are obtained through the prediction result refinement module, which is the final rainfall prediction result.

[0089] In one embodiment, the pre-training process of the advection simulator or the near-falling rainfall prediction network model is evaluated by MSE loss, and the Adam algorithm is used to minimize the MSE loss to obtain the pre-trained advection simulator or the near-falling rainfall prediction network model, and the MSE loss is as follows:

[0090]

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

[0092] In the training process of the advection simulator or the near-falling rainfall prediction network model, the data set used is a number of radar echo images, each image has a pixel of 288×288, and the time resolution of the image is 5 minutes, 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 data set randomly divides all radar echo images, and obtains 4000 sequences in the training set, 1557 sequences in the test set, and 1734 sequences in the validation set. For example, the near-falling rainfall prediction network model training is carried out on the training set, and the network model adopts random values ​​for the initial weight value, and the learning rate is 0.0001. In this embodiment, a total of 100 complete trainings are carried out. And each training is carried out once, and the loss value is verified on the validation set. The loss value is obtained by calculating the MSE loss of the final rainfall prediction result and the last q frame data (real rainfall data, i.e., the true value) of the rainfall image sequence in the validation set, and the Adam algorithm is used to minimize the loss, and finally the pre-trained near-falling rainfall prediction network model is obtained.

[0093] In one embodiment, the pre-trained advection simulator or the pre-trained near-falling rainfall prediction network model is further subjected to comparative testing using two binary indices and two non-binary indices, wherein:

[0094] The binary indicators include the critical success index CSI and the Heidecke skill score HSS, and the formula is 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-binarization indicators include balanced mean square error B-MSE and balanced mean absolute error B-MAE, the formulas are as follows:

[0098]

[0099] in,

[0100]

[0101] In the formula, w t,i,j represents the balance weight of the pixel (i, j) in the t-th frame rainfall image, w(x) represents the balance weight of a single pixel in the rainfall image, x t,i,j represents the rainfall at pixel (i, j) of the t-th frame rainfall image, represents the rainfall at 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 results and the corresponding real rainfall data is mainly through two binary indicators and two non-binary indicators. The binary indicators will compare the two according to the set rainfall intensity threshold, and the non-binary indicators are used to evaluate the similarity between the two from another perspective.

[0103] The binarization indicators used are the Critical Success Index (CSI) and the Heidke Skill Score (HSS). The CSI measures the ratio of the number of correctly predicted samples to the sum of all correctly and incorrectly predicted samples, and the HSS measures the improvement of the prediction compared to random prediction. The binarization indicators are mainly calculated by the following four indicators: 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 to moderate rain, moderate rain, and heavy rain and above, corresponding to the rainfall amounts of 0.5 mm / h, 2 mm / h, 5 mm / h, and 10 mm / h as thresholds, so as to perform binary division between the predicted results and the actual rainfall conditions. For the CSI and HSS indicators, the larger the value of the indicator, the more similar the predicted result is to the actual value.

[0105] Since there is an extremely unbalanced data problem in the data set, specifically, the data without rain in the data set is much more than the data set with rain, and the number of heavy rains is also less than the number of light rains. Therefore, the non-binary indicators Balanced Mean Square Error (B-MSE) and Balanced Mean Absolute Error (B-MAE) used introduce balanced weights, and assign different weights according to the rainfall intensity. For the non-binary indicators B-MSE and B-MAE, the lower the indicator value, the smaller the error between the predicted result and the actual rainfall situation, and the better the prediction effect, that is, the closer the predicted result is to the actual value.

[0106] The advection simulator was trained separately, using B-MAE as the loss function, using random values ​​for the initial weights, and a learning rate of 0.0001. In this embodiment, a total of 100 complete trainings were performed. The prediction results obtained are Figure 4 It is shown in the figure that the advection simulator has learned the overall direction of rainfall movement and the local intensity changes. 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 the figure. Figure 5 As shown in the figure, the advection prediction result can better predict rainfall than the evolution prediction result, and after the prediction time increases, the predicted rainfall area will not spread and the predicted image will not be blurred. Where T refers to the starting time of the prediction, that is, the time point of the 6th frame of rainfall image.

[0107] The near-term rainfall prediction network model (PiCNet) in this method is compared with the existing models with better performance, ConvLSTM (Convolutional Long Short-Term Memory Network), PredRNN (Recurrent Neural Network for Spatiotemporal Predictive Learning), PredRNN++ (an enhanced version of PredRNN), PFST, and MIM (Memory in memory learning high-order nonlinear spatiotemporal dynamic neural network) by four evaluation indicators, as shown in Table 1:

[0108] Table 1

[0109]

[0110]

[0111] According to Table 1, PiCNet outperforms the existing models in terms of HSS, CSI, and B-MSE for light rain and above rainfall intensity, and its B-MAE is basically the same as that of the existing models, especially in the heavy rainfall part (x ≥ 10), HSS and CSI have been significantly improved compared with other models, and the overall prediction accuracy has been significantly improved.

[0112] The prediction results of PiCNet, ConvLSTM, PredRNN, PredRNN++ and PFST for the rainfall in the next five minutes are Figure 6 Compared with the actual value of the rainfall in the last five minutes, Groundtruth, it can be seen that PiCNet's prediction results are better than other models, especially in areas with heavy rainfall, where more accurate prediction results can be obtained.

[0113] Among them, the evolution prediction method in NowcastNet refers to the technical solution adopted by 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 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 inApplied Earth Observations and Remote Sensing, 2020, 14:843-857.

[0114] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0115] The above-described embodiments only express the more specific and detailed embodiments described in this application, but they cannot be understood as limiting the scope of the patent application. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of this application, which all belong to the protection scope of this application. Therefore, the protection scope of the patent application shall be based on the attached claims.

Claims

1. A near-term rainfall prediction method based on physical constraints, characterized by: The near-term rainfall prediction method based on physical constraints comprises the following steps: S1, preprocessing each frame image in the meteorological radar echo sequence to obtain a corresponding rainfall image sequence; S2, based on the pre-trained advection simulator, the first p frames of rainfall images in the rainfall image sequence are used to predict the last q frames of rainfall images as the advection prediction result, where p+q=N, N is the total number of rainfall image frames in the rainfall image sequence; S3. The advection prediction result is combined with the previous p frames of rainfall image as a physical constraint and input into a pre-trained near-term rainfall prediction network model to obtain the final rainfall prediction result. The near-term rainfall prediction network model includes a physical guidance prediction module and a prediction result refinement module, wherein: The physical guidance prediction module includes an advection prediction result feature extraction module and a gate mechanism module, wherein the advection prediction result feature extraction module is used to extract the maximum value and the average value of the advection prediction result respectively through a first calculation module and a second calculation module in parallel, and add the maximum value and the average value to obtain a first extracted feature, wherein the first calculation module includes a global maximum 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; The gate mechanism module is used to splice the rainfall image of the previous p frames and the advection prediction result to obtain a splicing feature, and convert the splicing feature into a second extraction feature of the same dimension as the advection prediction result through a convolution layer, and then multiply the second extraction feature with the first extraction feature through a Sigmoid activation layer to obtain a third extraction feature, and finally add the second extraction feature to the third extraction feature to obtain a 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 encoders and second decoders in the same layer are connected correspondingly, the output features of the second encoders in each layer and the output features of the second decoders in the next layer are used as the input features of the second decoders in the current layer, the output features of the second encoders in the last layer and themselves are used as the input features of the second decoders in the current layer, the rainfall images of the previous p frames, the advection prediction results and the physical guidance prediction results are spliced ​​as the input features of the second encoders in the first layer, and the output features of the second decoders in the first layer are the final rainfall prediction results.

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

3. The near-term rainfall prediction method based on physical constraints as claimed in claim 1, characterized in that: The advection simulator performs the following operations: S2.1, inputting the first p frames of rainfall images of the rainfall image sequence into the first encoder to obtain the fourth extracted features, and then inputting the fourth extracted features into two parallel first decoders to output the velocity field and rainfall residual field of the next q frames of rainfall images respectively; S2.2, inputting the velocity fields of the p-th frame rainfall image and the p+1-th frame rainfall image into the Euler integrator, and adding 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; S2.3, set p=p+1, and return to execute step S2.2 until the next q frames of rainfall images are obtained, and the next q frames of rainfall images are used as advection prediction results.

4. The near-term rainfall prediction method based on physical constraints as claimed in claim 3, characterized in that: The Euler integrator adopts the 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; Where Langrangian represents the semi-Lagrangian integral, X represents the rainfall image of the current frame, 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 integral calculation, X2 represents the estimated rainfall image of the current frame obtained by reverse semi-Lagrangian integral calculation of X1, 2e represents the error between the forward semi-Lagrangian integral and the reverse semi-Lagrangian integral, that is, the error sum of the two semi-Lagrangian integrals, It represents the estimated rainfall image after deducting the error Δt, that is, the output result of the Euler integrator.

5. The near-term rainfall prediction method based on physical constraints as claimed in claim 1, characterized in that: The pre-training process of the advection simulator or the near-falling rainfall prediction network model 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 the near-falling rainfall prediction network model. The MSE loss is as follows: In the formula, MSE represents the MSE loss, y t represents the actual rainfall result of the t-th frame rainfall image, that is, the t-th frame rainfall image in the rainfall image sequence, It represents the t-th frame rainfall image output by the advection simulator or the near-fall rainfall prediction network model.

6. The near-term rainfall prediction method based on physical constraints as claimed in claim 5, characterized in that: The pre-trained advection simulator or the pre-trained near-falling rainfall prediction network model is also subjected to comparative tests using two binary indices and two non-binary indices, wherein: The binary index includes the critical success index CSI and the Heidecke skill score HSS, and the formula is as follows: 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; The non-binarization index includes the balanced mean square error B-MSE and the balanced mean absolute error B-MAE, and the formula is as follows: in, In the formula, w t,i,j represents the balance weight of the pixel (i, j) in the t-th frame rainfall image, w(x) represents the balance weight of a single pixel in the rainfall image, x t,i,j represents the rainfall at pixel (i, j) of the t-th frame rainfall image, represents the rainfall at 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.

7. The near-term rainfall prediction method based on physical constraints according to any one of claims 1 to 6, characterized in that: Each of the encoders includes five first feature extraction modules connected in sequence, and the first feature extraction modules include a maximum pooling layer, a convolution layer, a normalization layer, a ReLU activation function, a convolution 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, and the second feature extraction modules include a bilinear interpolation layer, a convolution layer, a normalization layer, a ReLU activation function, a convolution layer, a normalization layer, and a ReLU activation function connected in sequence.

Images