Method, computer device, system, computer readable storage medium and computer program product for weather forecasting

Abstract

The present invention relates to the field of computer systems using computational models, and specifically provides methods, computer devices, systems, computer-readable storage media, and computer program products for weather forecasting. A method for weather forecasting includes: obtaining weather data; extracting deep learning features from the weather data; estimating the horizontal displacement corresponding to the movement of a reference unit in a first frame from a reference position to a target position in a second frame in a feature space; determining the vertical displacement of the reference unit in the feature space according to a three-dimensional continuity equation based on the horizontal displacement; performing directional convolution operations on the deep learning features; and weighting the directional convolution results based on the vertical displacement. The present invention determines the vertical displacement in the feature space through a three-dimensional continuity equation, and selects the direction of the deep learning features based on the vertical displacement, combines the state of the atmosphere in the vertical direction in the real space into the calculation of the deep learning network, and increases the accuracy of weather forecasting.

Description

Technical Field

[0001] The present invention generally relates to computer systems using computational models, and more particularly to methods, computer devices and systems for weather forecasting. Background Art

[0002] There is extensive research in the field of meteorological forecasting, especially nowcasting.

[0003] Chinese patent application publication CN 108693534 A discloses an NRIET X-band radar collaborative networking analysis method, which uses reflectivity data of multiple layers (e.g., multiple planes at a certain height from a reference plane) detected by multiple radars, inverts the (u, v) components in the three-dimensional wind field through radial velocity, and then solves the vertical component w through the three-dimensional continuity equation to correct the original multi-layer reflectivity data detected by the radar.

[0004] In recent years, deep learning network technology has also been widely used in weather forecasting.

[0005] Some techniques utilize deep learning networks to perform weather forecasting based on detected weather data (e.g., reflectivity data detected by radar) of a single layer (e.g., a single plane at a certain height from a reference plane).

[0006] Some techniques use autoregressive prediction methods. For example, some methods use convLSTM to extract the spatial structure information of meteorological data detected by radar through convolution operations, and then use the Long Short Term Memory network to establish the relationship between different frames in the time dimension, so as to predict future meteorological data (see https: / / proceedings.neurips.cc / paper / 5955-convolutional-lstm-network-a-machine-learningapproach-for-precipitation-nowcasting.pdf). Some methods use PredRNN for autoregressive prediction (see https: / / arxiv.org / abs / 2103.09504).

[0007] Some technologies use non-autoregressive (multiple-input, multiple-output) prediction methods. For example, some methods use SimVP to input multiple frames of radar images to the network at the same time, and then predict multiple frames of images in the future at the same time. Among them, the network structure often uses the U-Net structure, which strengthens the feature representation function through network upsampling, downsampling, lateral connection, etc.

[0008] Some techniques use deep learning networks to make weather forecasts based on multiple layers of radar-detected meteorological data (e.g., reflectivity data).

[0009] Some technologies use deep learning networks to make weather forecasts based on multi-layer radar detection meteorological data. Some methods use reflectivity data on multi-layer radar echo images to perform separate feature extraction on high-level reflectivity data, and then superimpose it on channels with low-level reflectivity data to enhance the expression ability of low-level reflectivity data for predicting future reflectivity data of the low level (see https: / / ieeexplore.ieee.org / document / 10538240).

[0010] There is a need in this area for techniques that better utilize data reflecting the true state of the atmosphere to produce more accurate weather forecasts. Summary of the invention

[0011] In order to provide a more accurate weather forecast solution, the present invention is provided.

[0012] One aspect of the present invention provides a method for weather forecasting, which includes: obtaining meteorological data; extracting deep learning features from the meteorological data, the deep learning features including multiple frames in a feature space, the feature space including a time dimension and a space dimension, the space dimension including a horizontal dimension in a plane at a predetermined height relative to a reference plane and a vertical dimension perpendicular to the reference plane; estimating in the feature space a horizontal displacement corresponding to a reference unit in a first frame of the multiple frames moving from a reference position in the first frame to a target position in a second frame of the multiple frames, wherein the second frame is adjacent to the first frame in the time dimension, and the horizontal displacement is associated with the horizontal dimension; based on the horizontal displacement, determining in the feature space according to a three-dimensional continuity equation a vertical displacement of the reference unit moving from the reference position to the target position, and the vertical displacement is associated with the vertical dimension; performing a directional convolution operation on the deep learning features to determine a directional convolution result; and weighting the directional convolution result based on the vertical displacement to determine a modified deep learning feature, and the modified deep learning feature is used to predict meteorological data at a future moment.

[0013] As in the method described above, the meteorological data is obtained by selecting raw meteorological data at preset height intervals between a preset initial height and an end height.

[0014] A method as described in any one of the above items, wherein the original meteorological data includes radar echo images or atmospheric variables obtained at different altitudes.

[0015] A method as described in any of the above items, wherein extracting deep learning features includes: performing downsampling on the meteorological data.

[0016] A method as described in any of the above items, wherein the horizontal dimension includes a first horizontal dimension and a second horizontal dimension, and estimating the horizontal displacement includes: estimating the first horizontal displacement, the first horizontal displacement is associated with the first horizontal dimension; and estimating the second horizontal displacement, the second horizontal displacement is associated with the second horizontal dimension.

[0017] A method as described in any of the above items, wherein estimating the horizontal displacement includes: calculating the feature similarity of the reference unit relative to the unit at each potential target position in the neighborhood in the second frame; using the feature similarity to weight the potential horizontal displacement of the reference unit moving from the reference position to each potential target position in the neighborhood; and estimating the weighted sum of the potential horizontal displacements associated with each potential target position as the horizontal displacement.

[0018] A method as described in any of the above items, wherein estimating the horizontal displacement comprises: performing optical flow deformation on the first frame after the reference unit is moved by a potential horizontal displacement; and estimating the horizontal displacement by minimizing the error between the second frame and the first frame after optical flow deformation.

[0019] A method as described in any of the above items, wherein determining the vertical displacement of the reference unit from the reference position to the target position includes: determining the vertical partial differential component of the reference unit from the reference position to the target position in the feature space according to a three-dimensional continuity equation based on the horizontal displacement; and integrating the vertical partial differential component in the vertical dimension to determine the vertical displacement.

[0020] A method as described in any of the above items, wherein the horizontal dimension includes a first horizontal dimension and a second horizontal dimension, and performing a directional convolution operation on the deep learning feature includes: decomposing the spatial dimension into a first plane defined by the first horizontal dimension and the vertical dimension and a second plane defined by the second horizontal dimension and the vertical dimension; performing a directional convolution operation on the deep learning feature in a first plurality of directions on the first plane to obtain a first plurality of directional convolution results; performing a directional convolution operation on the deep learning feature in a second plurality of directions on the second plane to obtain a second plurality of directional convolution results; and combining the first plurality of directional convolution results and the second plurality of directional convolution results to determine the directional convolution result.

[0021] A method as described in any of the above items, wherein weighting the directional convolution result includes: determining a first plurality of weights and a second plurality of weights based on the vertical displacement, the first plurality of weights being associated with the first plurality of directions, and the second plurality of weights being associated with the second plurality of directions; and weighting the directional convolution result with the first plurality of weights and the second plurality of weights.

[0022] A method as described in any of the above items, wherein weighting the directional convolution result includes: determining a first plurality of weights and a second plurality of weights based on the vertical displacement and the horizontal displacement, the first plurality of weights being associated with the first plurality of directions, and the second plurality of weights being associated with the second plurality of directions; and weighting the directional convolution result with the first plurality of weights and the second plurality of weights.

[0023] A method as described in any of the above items, wherein the method further comprises: performing feature extraction on the corrected deep learning features to determine predicted deep learning features; and performing weather forecasting based on the predicted deep learning features.

[0024] Another aspect of the present invention provides a computer device for weather forecasting, comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement any of the above methods.

[0025] Another aspect of the present invention provides a system for weather forecasting, the system comprising: a computer device as described above; and a detection device coupled to the computer device and used to detect the weather data.

[0026] Another aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the above methods.

[0027] Another aspect of the present invention provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the steps of any one of the above methods are implemented.

[0028] The method, computer device and system according to the embodiments of the present invention can achieve the following technical effects:

[0029] Compared with the traditional non-deep learning method that requires data detected by multiple radars for weather forecasting, some embodiments can predict weather data at future times through a deep learning network based on multiple layers of weather data detected by a single radar (e.g., radar echo images and associated reflectivity data) to perform weather forecasting. In addition, some embodiments perform relevant calculations and estimates in feature space rather than real space (original pixel space), so that the accuracy of relevant calculations and estimates can be improved as the quality of features is improved, and the range of applicable input weather data is wider.

[0030] Compared with the prior art that performs weather forecasting based on only a single layer of weather data, some embodiments utilize multiple layers of weather data for forecasting, which can better characterize the actual atmospheric state, thereby performing weather forecasting more accurately.

[0031] Compared with the existing technology of using deep learning networks and multi-layer meteorological data to make weather forecasts, some embodiments introduce data interaction at different heights in the vertical dimension, calculate the corresponding vertical displacement by horizontal displacement from the reference layer to different target layers, and incorporate the physical laws of changes in the vertical dimension into the calculation process of the deep learning network, so as to combine the mutual interaction and influence of the atmospheric state in the real space in the vertical dimension into the weather forecast at the future moment, making the forecast result more accurate. In addition, some embodiments guide the directional convolution results of deep learning features based on vertical displacement, and can select the potential directional features of the directional convolution results to the greatest extent, thereby characterizing the real movement direction and state of the atmosphere to the greatest extent.

[0032] Experimental results show that, compared with the method of predicting radar echoes based on single-layer radar transmission, the embodiments of the present invention can improve the critical success index (CSI) by nearly 1.5 percentage points based on the baseline model. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Embodiments of the present invention are described with reference to the accompanying drawings.

[0034] Figure 1 A flow chart showing a method for weather forecasting according to some embodiments of the present invention.

[0035] Figure 2 A first flow chart illustrating a first process of estimating horizontal displacement according to some embodiments of the present invention.

[0036] Figure 3 A second flow chart illustrating a second process for estimating horizontal displacement according to some embodiments of the present invention.

[0037] Figure 4is a schematic diagram of estimating possible horizontal displacement within a neighborhood of a reference unit in one frame moving to the next frame according to some embodiments of the present invention.

[0038] Figure 5 is a schematic diagram of a process of estimating horizontal displacement according to some embodiments of the present invention.

[0039] Figure 6 A third flow chart illustrating a third process for estimating horizontal displacement according to some embodiments of the present invention.

[0040] Figure 7 A flow chart illustrating a fourth process for determining a vertical displacement of a reference unit moving from a reference position to a target position according to some embodiments of the present invention.

[0041] Figure 8 is a schematic diagram of a reference unit in one frame moving from a reference position to a target position in a neighborhood in an adjacent frame according to some embodiments of the present invention.

[0042] Fig. 9 A flowchart illustrating a fifth process of performing directional convolution operations on deep learning features according to some embodiments of the present invention.

[0043] Fig.10 A first flow chart of a sixth process for weighting directional convolution results according to some embodiments of the present invention is shown.

[0044] Fig.11 A second flow chart of a seventh process for weighting directional convolution results according to some embodiments of the present invention is shown.

[0045] Fig.12 A flowchart of an eighth process of performing directional convolution operations on deep learning features and weighting directional convolution results according to some embodiments of the present invention is shown.

[0046] Fig.13 is a schematic diagram of a ninth process of performing directional convolution operations on deep learning features and weighting directional convolution results according to some embodiments of the present invention.

[0047] Fig.14 is a schematic diagram of a convolution kernel used for directional convolution operation according to some embodiments of the present invention.

[0048] Fig.15 A block diagram illustrating a computer device for weather forecasting according to some embodiments of the present invention.

[0049] Fig.16 A block diagram of a system for weather forecasting according to some embodiments of the present invention is shown.

[0050] Fig.17 A block diagram illustrating a computer-readable storage medium according to some embodiments of the invention.

[0051] Fig.18 A block diagram illustrating a computer program product according to some embodiments of the invention. DETAILED DESCRIPTION

[0052] In the present application, the term "reference plane" refers to a plane that can be used as a reference for meteorological detection, such as the ground, sea level, a reference plane with an altitude of zero, etc. The term "horizontal dimension" refers to a dimension in a plane at a certain height relative to the reference plane, for example, a dimension in a plane of equal height relative to the reference plane. The term "horizontal displacement" refers to the displacement of a reference unit when it moves in the horizontal dimension. The term "vertical dimension" refers to a dimension perpendicular to the reference plane, for example, the height dimension. The term "vertical displacement" refers to the displacement of a reference unit when it moves in the vertical dimension. The term "layer" refers to a plane relative to the "reference plane" in the "vertical dimension". The term "single layer" means a plane corresponding to one height, and the term "multi-layer" means a plane corresponding to multiple different heights.

[0053] According to some embodiments of the present invention, a method for weather forecasting is provided.

[0054] Figure 1 A flow chart of a method 100 for weather forecasting according to some embodiments of the present invention is shown.

[0055] Method 100 may include step 110: obtaining meteorological data.

[0056] Meteorological data can be data that can be obtained by means of meteorological detection and can be used for weather forecasting. In some embodiments, meteorological data can be obtained based on raw meteorological data detected by a detector. In some embodiments, meteorological data can be obtained by selecting raw meteorological data at a preset altitude interval between a preset initial altitude and an end altitude.

[0057] In some embodiments, the raw meteorological data may include radar echo images obtained at different heights relative to a reference plane. The radar echo image may be detected by a meteorological radar. The radar echo image may include reflectivity data. The reflectivity data may reflect the concentration of particles in the atmosphere within a certain volume in space. As an example, the raw meteorological data may include radar echo images (and associated reflectivity data) at 0-21 km. As an example, the initial height may be preset to 2 km, the end height to 11 km, and the height interval to 1.5 km. Then, radar echo images at heights of 2 km, 3.5 km, 5 km, 6.5 km, 8 km, 9.5 km, and 11 km may be selected from the raw radar echo images obtained at 0-21 km as inputs for further processing.

[0058] In some embodiments, the meteorological data may include atmospheric variables. As an example, the atmospheric variables may include atmospheric pressure, atmospheric temperature, etc. The atmospheric variables may be detected by corresponding detectors or sensors.

[0059] The method 100 may include step 120: extracting deep learning features from the meteorological data. The deep learning features may include multiple frames in a feature space. The feature space may include a time dimension and a space dimension. The space dimension may include a horizontal dimension in a plane at a predetermined height relative to a reference plane and a vertical dimension perpendicular to the reference plane.

[0060] In some embodiments, various network frameworks for deep learning can be used to extract deep learning features. As an example, deep learning networks such as ResNet and Transformer can be used.

[0061] As an example, the dimension of the feature space can be [T, Z, C, X, Y], where T is the time dimension, which may correspond to the number of frames of meteorological data obtained, Z is the vertical dimension, which may correspond to the number of layers corresponding to the meteorological data, C is the number of channels of the deep learning feature, which is used in the deep learning operation process, and X and Y are horizontal dimensions, which may correspond to the sizes of the deep learning features in two dimensions in the plane.

[0062] In some embodiments, step 120 may include: downsampling the meteorological data. As an example, the meteorological data may be downsampled by a factor of S to obtain the dimensions X and Y of the features in two dimensions.

[0063] The method 100 may include step 130: estimating in a feature space a horizontal displacement corresponding to a reference unit in a first frame of the plurality of frames moving from a reference position in the first frame to a target position in a second frame of the plurality of frames. The second frame may be adjacent in a temporal dimension relative to the first frame. The horizontal displacement may be associated with a horizontal dimension. The second frame may be different or the same in a vertical dimension relative to the first frame.

[0064] As an example, the dimensions of the first frame and the second frame may be [T, Z, C, X, Y] as described above. The first frame may be represented as F1(t, z), and the second frame may be represented as F2(t+Δt, z+Δz).

[0065] The following will combine Figure 2-Figure 6 The process of determining the horizontal displacement is further described.

[0066] Method 100 may include step 140: determining a vertical displacement of the reference unit from the reference position to the target position according to a three-dimensional continuity equation in the feature space based on the horizontal displacement. The vertical displacement may be associated with a vertical dimension. Figure 7 , Figure 8 The process of determining the vertical displacement is further described.

[0067] The method 100 may include step 150: performing a directional convolution operation on the deep learning feature to determine a directional convolution result. Figure 9-11 The process of performing a directional convolution operation is further described.

[0068] The method 100 may include step 160: weighting the directional convolution result based on the vertical displacement to determine a modified deep learning feature. The modified deep learning feature can be used to predict weather data at a future time. Fig.12 , Fig.13 The process of performing directional convolution operations on deep learning features and weighting the directional convolution results based on vertical displacement is further described. Based on the predicted weather data at a future time, weather forecasts can be performed.

[0069] Some embodiments extract deep learning features from meteorological data, perform directional convolution operations on the deep learning features, and estimate horizontal and vertical displacements based on the extracted deep learning features. The results of the directional convolution are weighted based on the estimated vertical displacement to adaptively select the convolution results in the best direction, thereby correcting the original deep learning features to predict the deep learning features at the next moment for weather forecasting.

[0070] Optionally, in some embodiments, method 100 may include step 170: performing feature extraction on the modified deep learning features to determine predicted deep learning features.

[0071] In some embodiments, various network frameworks for deep learning can be used to extract features. As an example, deep learning networks such as ResNet and Transformer can be used.

[0072] In some embodiments, upsampling may be performed on the extracted features to determine predicted deep learning features.

[0073] Optionally, in some embodiments, method 100 may include step 180: performing weather forecasting based on predicted deep learning features.

[0074] Figure 2 A first flow chart illustrating a first process 200 for estimating horizontal displacement according to some embodiments of the present invention is shown. The first process 200 may be Figure 1 The specific implementation of step 130 of method 100 is described below, but the scope of the present invention is not limited thereto.

[0075] In some embodiments, the horizontal dimension may include a first horizontal dimension and a second horizontal dimension.

[0076] The first process 200 may include a step 210 of estimating a first horizontal displacement. The first horizontal displacement may be associated with a first horizontal dimension.

[0077] The first process 200 may include a step 230 of estimating a second horizontal displacement. The second horizontal displacement may be associated with a second horizontal dimension.

[0078] In some embodiments, the direction corresponding to the first horizontal dimension may be perpendicular to the direction corresponding to the second horizontal dimension. Some embodiments may combine the two-dimensional convolution operation in the plane by estimating the corresponding horizontal displacement in two perpendicular directions in the plane, thereby facilitating calculation using a deep neural network.

[0079] In some embodiments, step 210 may be performed in parallel with step 230. In some embodiments, step 210 may be performed in series with step 230. As an example, step 210 may be performed first, and then step 230. As an example, step 230 may be performed first, and then step 210.

[0080] The first process 200 estimates the horizontal displacement of the reference unit moving from a reference position to a target position in two different horizontal dimensions.

[0081] Figure 3 A second flow chart illustrating a second process 300 for estimating horizontal displacement according to some embodiments of the present invention is shown. The second process 300 may be Figure 1 The specific implementation of step 130 of method 100 is described below, but the scope of the present invention is not limited thereto.

[0082] The second process 300 may include step 310 of computing feature similarities of a reference cell relative to cells at each potential target location in a neighborhood in a second frame.

[0083] The second process 300 may include step 330 of weighting a potential horizontal displacement of a reference unit moving from a reference location to each potential target location in a neighborhood using feature similarity.

[0084] The second process 300 may include a step 350 of estimating a weighted sum of potential horizontal displacements associated with each potential target location as a horizontal displacement.

[0085] The second process 300 determines a potential horizontal displacement based on the reference position of the reference unit to each potential target position in the neighborhood, and uses the feature similarity between the reference frame and the unit at each potential target position as a weight to weight the potential horizontal displacement to determine the horizontal displacement. The feature similarity can be used to characterize the probability of the reference unit moving to the potential target position in the neighborhood.

[0086] In some embodiments, the second process 300 may be combined with Figure 2 The first process 200 in FIG. 4 is performed. That is, the feature similarity between the reference unit and the unit at each potential target position in the neighborhood is determined, and the two potential horizontal displacements on two different horizontal dimensions are weighted and summed by the feature similarity to obtain two horizontal displacements respectively associated with the two different horizontal dimensions.

[0087] The following will combine Figure 4 , Figure 5 Further details of performing the second process 300 are further described.

[0088] Figure 4 is a schematic diagram of estimating possible horizontal displacement within a neighborhood of a reference unit in one frame moving to the next frame according to some embodiments of the present invention.

[0089] Figure 4 A first frame 410 and a second frame 450 are shown in FIG. 4 , and two horizontal dimensions X, Y are shown that are perpendicular to each other. Figure 4 The direction perpendicular to the paper plane may correspond to the vertical dimension. The first frame 410 may be represented as (t, z), which may correspond to a frame at a previous time t and a height z in the vertical dimension. The second frame 450 may be represented as (t+Δt, z+Δz), which may correspond to a frame at a later time t+Δt and a height z+Δz in the vertical dimension. In some embodiments, Δz may be zero.

[0090] The first frame 410 may include a first reference unit 415. The first reference unit 415 may be represented as (t, z, x, y), which may indicate that the first reference unit 415 is located in the first frame 410 at a time t, at a height z in the vertical dimension, and has a plane coordinate (x, y) in the horizontal dimension.

[0091] In order to determine the area to which the first reference unit 415 may move at the next moment t+Δt, the same position as the first reference unit 415 in the first frame 410 may be used as a reference position in the second frame 450, and a distance Δp may be extended on both sides in the directions of the dimensions X and Y to form a first neighborhood 455 in the second frame 450. As an example, the maximum distance of movement on both sides in the directions of the dimensions X and Y is shown to be P. Therefore, the first reference unit 415 in the second frame 450 may be represented as (t+Δt, z+Δz, h+Δp, w+Δp), where Δp∈[-P, P].

[0092] In some embodiments, the relevant features C(t, z, x, y) may be determined for the first reference cell 415 in the first frame 410 and the cell at each potential target position in the first neighborhood 455 in the second frame 450 according to the following formula (1):

[0093] C(t, z, x, y)={Sim (F(t, z, x, y), F(t+△t, z+△z, x+△p, y+△p)|△p∈[-P, P])} (1)

[0094] Wherein, Sim() represents a similarity function, F(t, z, x, y) represents a feature of a first reference cell 415 in the first frame 410 , and F(t+Δt, z+Δz, x+Δp, y+Δp) represents a feature of a cell at a potential target position in the second frame 450 .

[0095] Therefore, the dimension of the input feature F can be [T, Z, C, X, Y], where T is the time dimension, Z is the vertical dimension, C is the channel dimension, and X and Y are two horizontal dimensions. The dimension of the correlation feature C can be [T, Z×Z, (2×P+1)×(2×P+1), X, Y]. That is, for the reference frame, the correlation value C(t+△t, z+△z) between the feature and the frame at each height on the vertical dimension Z at the next moment can be calculated; for the reference unit (t, z, x, y) in the reference frame, the feature similarity can be calculated between the unit (t+△t, z+△z, x+△p, y+△p) at each potential target position in the neighborhood of the reference unit in the frame at each height on the vertical dimension Z at the next moment. This feature similarity can be used to characterize the probability of each reference unit moving to the potential target position in the corresponding neighborhood.

[0096] The relevant feature C(t, z, x, y) calculated for a specific reference unit (t, z, x, y) can be marked as ,in, In some embodiments, the obtained relevant features C can be normalized into a probability vector whose sum is one by a softmax function. Subsequently, the horizontal displacements U and V in the directions corresponding to the two horizontal dimensions can be calculated by multiplying the probability of the potential horizontal displacement and the corresponding displacement by the following formulas (2) and (3):

[0097] (2)

[0098] (3)

[0099] Combined with the following Figure 5 Describe the process of using a deep learning network to calculate horizontal displacement.

[0100] Figure 5 is a schematic diagram of a process of estimating horizontal displacement according to some embodiments of the present invention. Figure 5 The process shown in can be Figure 1 Specific implementation of step 130 of method 100, Figure 2 A specific implementation of the first process 200, Figure 3 A specific implementation of the second process 300 in Figure 4 The specific implementation manner of the process in the embodiment of the present invention is described below, but the scope of the present invention is not limited thereto.

[0101] The second reference unit 510 may be a unit in a previous frame (eg, Figure 4 The second neighborhood 520 may be a corresponding neighborhood in a subsequent frame (eg, Figure 4In the illustrated embodiment, the second neighborhood 520 is divided into 9 potential target locations, each of which may correspond to a unit, such as Figure 5 The feature similarity of the second reference unit 510 relative to the unit in each potential target position in the second neighborhood 520 can be calculated to obtain a feature similarity matrix 530. The number shown at the position corresponding to each potential target position in the feature similarity matrix 530 is the feature similarity of the corresponding potential target unit.

[0102] The potential horizontal displacement of the second reference cell 510 in the second neighborhood 520 may be determined ( Figure 5 The plane where the paper is located can correspond to the horizontal dimension, so the horizontal displacement can correspond to the displacement in this horizontal dimension). The center position of the nine-square grid formed by the second neighborhood 520 is used as the starting point of the movement of the second reference unit 510, and the potential horizontal displacement corresponding to its movement to the adjacent position is determined respectively. In the illustrated embodiment, the corresponding potential horizontal displacement is determined in two mutually perpendicular horizontal dimensions. As shown in the potential horizontal displacement matrix 540, the center position of the nine-square grid is used as the reference position, and its position is marked as (0, 0), which indicates that the potential horizontal displacement in the two horizontal dimensions is 0. For the position to the left of the reference position, the potential horizontal displacement of the reference unit moving to this position can be marked as (0, -1), which indicates that in the first horizontal dimension ( Figure 5 The potential horizontal displacement in the up and down directions in the second horizontal dimension ( Figure 5 The potential horizontal displacement of the reference unit to the left and right directions (in the left and right directions) is -1 (one position to the left). Similarly, for a position to the right of the reference position, the potential horizontal displacement of the reference unit to that position can be labeled (0, 1). Similarly, for a position above the reference position, the potential horizontal displacement of the reference unit to that position can be labeled (-1, 0). Similarly, for a position below the reference position, the potential horizontal displacement of the reference unit to that position can be labeled (1, 0). The potential horizontal displacement of the reference unit to other positions can also be determined and labeled in a similar manner.

[0103] The similarity at each position in the similarity matrix 530 may be used as a weight to weight the potential horizontal displacement at the corresponding position of the potential horizontal displacement matrix 540 to determine a weighted potential horizontal displacement matrix 550 .

[0104] As an example, for the position of the upper left corner of the weighted potential horizontal displacement matrix 550, the potential horizontal displacement (-1, -1) of the corresponding position (upper left corner) of the potential horizontal displacement matrix 540 can be weighted by the similarity 0.03 of the corresponding position (upper left corner) in the similarity matrix 530 to obtain the weighted potential horizontal displacement (-0.03, -0.03). The weighted horizontal displacements of the weighted potential horizontal displacement matrix 550 at other positions can be calculated in a similar manner.

[0105] The weighted potential horizontal displacements of weighted potential horizontal displacement matrix 550 may be summed in the first horizontal dimension and the second horizontal dimension, respectively, to determine horizontal displacement 560 .

[0106] As an example, the weighted potential horizontal displacements of the weighted potential horizontal displacement matrix 550 on the first horizontal dimension may be summed as follows: -0.03+(-0.04)+(-0.05)+0+0+0+0.04+0.20+0.20=0.32. As an example, the weighted potential horizontal displacements of the weighted potential horizontal displacement matrix 550 on the second horizontal dimension may be summed as follows: -0.03+0+(-0.05)+(-0.04)+0+0.20+0.04+0+0.20=0.32. Thus, the horizontal displacement 560 is determined to be (0.32, 0.32).

[0107] In some embodiments, Figure 5 The calculation process described in the above can be implemented through a Transformer deep learning network, wherein the second reference unit 510 can be used as a query (Query, Q) value, and each target unit in the second neighborhood 520 can be used as a key (Key, K) value.

[0108] Figure 6 A third flow chart illustrating a third process 600 for estimating horizontal displacement according to some embodiments of the present invention. The third process 600 may be Figure 1 The specific implementation of step 130 of method 100 is as follows, but the scope of the present invention is not limited thereto. The third process 600 estimates the horizontal displacement by optical flow deformation method.

[0109] The third process 600 may include step 610 of performing optical flow deformation on the first frame after the reference unit has been moved by a potential horizontal displacement.

[0110] The third process 600 may include step 630 of estimating a horizontal displacement by minimizing an error between the second frame and the first frame deformed by optical flow.

[0111] Figure 7Flowchart showing a fourth process 700 for determining a vertical displacement of a reference unit moving from a reference position to a target position according to some embodiments of the present invention. The fourth process 700 may be Figure 1 The specific implementation of step 140 in the embodiment of the present invention is described below, but the scope of the present invention is not limited thereto.

[0112] The fourth process 700 may include step 710 of determining a vertical partial differential component of a reference unit moving from a reference position to a target position according to a three-dimensional continuity equation in a feature space based on the horizontal displacement.

[0113] The fourth process 700 may include a step 730 of integrating the vertical partial differential component in the vertical dimension to determine a vertical displacement.

[0114] Combined with the following Figure 8 The specific implementation of the fourth process 700 is further described.

[0115] Figure 8 is a schematic diagram of a reference unit in one frame moving from a reference position to a target position in a neighborhood in an adjacent frame according to some embodiments of the present invention.

[0116] Figure 8 A third frame 810 at (t, z) in feature space and a fourth frame 820 at (t+Δt, z+Δz) are shown. In some embodiments, Δz may be zero.

[0117] The third frame 810 may include a third reference unit 815. The fourth frame 820 may include a third neighborhood 823 and a target position 825 in the third neighborhood 823. The third reference unit 815 may move to the target position 825 in the third neighborhood 823 in the fourth frame 820. The target position 825 in the third neighborhood 823 may be obtained by combining the above Figure 2-Figure 6 Describe the process to determine.

[0118] For such movement, the horizontal displacement of the third reference unit 815 in the horizontal dimension may be marked as U, V, and the vertical displacement in the vertical dimension may be marked as W. In some embodiments, U and V may be perpendicular to each other, and U, V, and W may correspond to movement in three directions defined by a three-dimensional Cartesian coordinate system.

[0119] The horizontal displacement of the reference unit in the horizontal dimension can be expressed by Figure 1 Step 130 of method 100 may be determined by Figure 2 The first process 200 determines that Figure 3 The second process 300 in the embodiment can be determined by Figure 4 The process shown in Figure 5 , which can be determined by the process shown in (eg, determined as a horizontal displacement 560) by Figure 6 The third process 600 in is determined, but the scope of the present invention is not limited in this regard.

[0120] According to the three-dimensional continuity equation of incompressible fluid in atmospheric dynamics shown in the following formula (4):

[0121] (4)

[0122] The vertical displacement W can be estimated based on the estimated horizontal displacements U and V. The three components in the three-dimensional continuity equation shown in formula (4) correspond to the flow velocity of the fluid in fluid mechanics, and the above-mentioned horizontal displacements U, V and vertical displacement W are the displacements of the reference unit from one frame to the target position on another frame at adjacent moments. These displacements divided by the time interval between adjacent frames can also determine the corresponding speed. On the other hand, deep learning features are extracted from meteorological data detected in real space. These meteorological data can reflect the state of the atmosphere, so the spatial change rate of meteorological data between adjacent moments in real space can follow the three-dimensional continuity equation. Correspondingly, the spatial change rate of deep learning features extracted from meteorological data between adjacent moments in feature space can also follow the three-dimensional continuity equation. Therefore, formula (4) is applicable to horizontal displacements U, V and vertical displacements W.

[0123] The partial differential component corresponding to the vertical direction can be calculated according to the following formula (5): :

[0124] (5)

[0125] Assume that the reference unit is (t, Z 1 , x, y), which is in the vertical dimension Z 1 The vertical displacement in the vertical dimension is W(t, Z 1 , x, y), the vertical displacement W(t, Z, x, y) of the unit (t, Z, x, y) can be calculated by the following formula (6):

[0126] (6)

[0127] The Z in the vertical dimension can be taken as the highest layer, where the vertical displacement is zero, and Z can be calculated according to the following formula (7): 1 Vertical displacement at:

[0128] (7)

[0129] Thus, for any reference unit (t, z, xy), the corresponding vertical displacement W(t, z, x, y) can be determined.

[0130] Fig. 9 FIG. 9 is a flowchart illustrating a fifth process 900 for performing a directional convolution operation on a deep learning feature according to some embodiments of the present invention. The fifth process 900 may be Figure 1 1 and 2. The present invention may be performed as step 150 of method 100 in the embodiment of the present invention, but the scope of the present invention is not limited in this regard.

[0131] The fifth process 900 may include step 910 of decomposing a spatial dimension into a first plane defined by a first horizontal dimension and a vertical dimension and a second plane defined by a second horizontal dimension and a vertical dimension.

[0132] The fifth process 900 may include step 930: performing directional convolution operations on the deep learning features in first multiple directions on the first plane, and performing directional convolution operations on the deep learning features in second multiple directions on the second plane. In some embodiments, the directional convolution operations performed in the first multiple directions may be performed in parallel with the directional convolution operations performed in the second multiple directions. In some embodiments, the directional convolution operations performed in the first multiple directions may be performed serially with the directional convolution operations performed in the second multiple directions. As an example, the directional convolution operations performed in the first multiple directions may be performed first, and then the directional convolution operations performed in the second multiple directions may be performed. As an example, the directional convolution operations performed in the second multiple directions may be performed first, and then the directional convolution operations performed in the first multiple directions may be performed. The deep learning features may be Figure 1 The deep learning features extracted in step 120 of method 100.

[0133] The fifth process 900 may include step 950 of combining the first plurality of directional convolution results and the second plurality of directional convolution results.

[0134] Some embodiments can extract the features of the deep learning features in a specific direction by performing directional convolution operations on the deep learning features on two planes respectively, thereby providing a basis for further utilizing vertical displacement to select the direction of the features.

[0135] The following will combine Fig.12 , Fig.13 The implementation details of the fifth process 900 are further described and will be combined with Fig.14 Describes the convolution kernel required to perform a directional convolution operation.

[0136] Fig.10 FIG. 1 is a first flow chart of a sixth process 1000 for weighting directional convolution results according to some embodiments of the present invention. The sixth process 1000 may be Figure 1 Specific implementation of step 150 of method 100, and can be combined with Fig. 9 The fifth process 900 is performed in the embodiment, but the scope of the present invention is not limited in this regard.

[0137] The sixth process 1000 may include step 1010 of determining a first plurality of weights and a second plurality of weights based on a vertical displacement.

[0138] The vertical displacement can be achieved by Figure 1 The vertical displacement determined in step 140 of method 100 or by combining the above Figure 6-Figure 8 The vertical displacement is determined by the description process.

[0139] A first plurality of weights may be associated with a first plurality of directions. The first plurality of directions may be Fig. 9 The first plurality of directions in step 930 of the fifth process 900 in the embodiment of the present invention. A second plurality of weights may be associated with a second plurality of directions, and the second plurality of directions may be Fig. 9 The second multiple directions in step 930 of the fifth process 900.

[0140] The sixth process 1000 may include step 1030 of weighting the directional convolution result with the first plurality of weights and the second plurality of weights. The directional convolution result may be Fig. 9 The directional convolution result obtained in step 950 of the fifth process 900.

[0141] Some embodiments determine multiple weights in directions corresponding to the corresponding directional convolution operations performed on the deep learning features based on the vertical displacement, and use these weights to weight the directional convolution results of the deep learning features, and use the vertical displacement to select deep learning features in specific directions, thereby introducing the movement and state of the atmosphere in the vertical dimension in the real space into the calculation of the deep learning network, thereby improving the accuracy of the deep learning feature prediction.

[0142] The following will combine Fig.12 , Fig.13 The implementation details of the sixth process 1000 are further described.

[0143] Fig.11 A second flow chart of a seventh process 1100 for weighting directional convolution results according to some embodiments of the present invention is shown. The seventh process 1100 may be Figure 1 Specific implementation of step 150 of method 100, and can be combined with Fig. 9 The fifth process 900 is performed in the embodiment, but the scope of the present invention is not limited in this regard.

[0144] The seventh process 1100 may include determining a first plurality of weights and a second plurality of weights based on the vertical displacement and the horizontal displacement at step 1110. The first plurality of weights may be associated with a first plurality of directions.

[0145] The vertical displacement can be achieved by Figure 1The vertical displacement determined in step 140 of method 100 or by combining the above Figure 6-Figure 8 The vertical displacement is determined by the process described in . The horizontal displacement can be determined by Figure 1 The horizontal displacement determined in step 130 of method 100 or by combining the above Figure 2-Figure 5 The description process determines the horizontal displacement.

[0146] A first plurality of weights may be associated with a first plurality of directions. The first plurality of directions may be Fig. 9 The first plurality of directions in step 930 of the fifth process 900 in the embodiment of the present invention. A second plurality of weights may be associated with a second plurality of directions, and the second plurality of directions may be Fig. 9 The second multiple directions in step 930 of the fifth process 900.

[0147] The seventh process 1100 may include step 1130: weighting the directional convolution result with the first plurality of weights and the second plurality of weights. The directional convolution result may be Fig. 9 The directional convolution result obtained in step 950 of the fifth process 900.

[0148] Some embodiments determine multiple weights in directions corresponding to the corresponding directional convolution operations performed on the deep learning features based on the vertical displacement and the horizontal displacement, and use these weights to weight the directional convolution results of the deep learning features, and use the vertical displacement and the horizontal displacement to select the deep learning features in a specific direction, thereby introducing the movement and state of the atmosphere in the vertical and horizontal dimensions in the real space into the calculation of the deep learning network, thereby improving the accuracy of the deep learning feature prediction.

[0149] The following will combine Fig.12 , Fig.13 Implementation details of the seventh process 1100 are further described.

[0150] Fig.12 FIG. 1 is a flowchart showing an eighth process 1200 for performing directional convolution operations on deep learning features and weighting directional convolution results according to some embodiments of the present invention. The eighth process 1200 may be Figure 1 The process of combining step 150 and step 160 of method 100, or Fig. 9 The fifth process 900 and Fig.10 The sixth process 1000 is performed in combination with the process, or Fig. 9 The fifth process 900 and Fig.11 The seventh process 1100 is performed in combination with the process.

[0151] First, a first deep learning feature is obtained 1210. The deep learning feature may have dimensions [T, X, C, X, Y], where T may correspond to the time dimension, Z may correspond to the vertical dimension, C may correspond to the channel dimension, and X and Y may correspond to the horizontal dimensions.

[0152] Subsequently, at step 1220, a plane decomposition may be performed on the first deep learning feature 1210 in the feature space to obtain a plane [Z, Y] 1222 and a plane [Z, X] 1224. The plane [Z, Y] 1222 may be defined by a vertical dimension Z and a horizontal dimension Y, and the plane [Z, X] 1224 may be defined by a vertical dimension Z and a horizontal dimension X. The features decomposed from the first deep learning feature 1210 on the plane [Z, Y] 1222 may have dimensions [T, X, C, Z, Y]. The features decomposed from the first deep learning feature 1210 on the plane [Z, X] 1224 may have dimensions [T, Y, C, Z, X].

[0153] Subsequently, at step 1226, directional convolution is performed on the first deep learning feature 1210 in N1 directions in the plane [Z, Y] 1222 to obtain directional convolution results. The directional convolution results may be dimensionally upgraded so that the obtained features have dimensions [T, X, N1×C, Z, Y]. Matrix operations are then performed on the features to obtain features [T, Z, N1×C, X,Y].

[0154] At step 1228, directional convolution is performed on the first deep learning feature 1210 in N2 directions in the plane [Z, X] 1224 to obtain directional convolution results. The directional convolution results may be dimensionally upgraded so that the obtained features have dimensions [T, Y, N2×C, Z, X]. Matrix operations are then performed on the features to obtain features [T, Z, N2×C, X, Y].

[0155] Subsequently, at step 1230, a plane combination is performed on the features obtained at steps 1226 and 1228 to obtain a directional convolution result 1240. The directional convolution result 1240 may have dimensions [T, Z, N1×N2, C, X, Y].

[0156] At step 1270, feature transformation may be performed on the vertical displacement W 1250 in directions corresponding to the N1 directions and the N2 directions to obtain weights 1275. The weights 1275 may have dimensions N1×N2. The vertical displacement W 1250 may be Figure 1 The vertical displacement determined in step 140 of method 100 or by combining the above Figure 7 , Figure 8In some embodiments, the vertical displacement W 1250 may be mapped to N1×N2 directions (i.e., the vertical dimension is mapped to N1×N2 dimensions) and normalized using a Softmax function to obtain N1×N2 vectors, which may be used to weight the directional convolution result 1240.

[0157] Optionally, in some embodiments, feature transformation may be performed on the vertical displacement W 1250 and the horizontal displacements U, V in directions corresponding to the N1 directions and the N2 directions to obtain weights 1275. The weights 1275 may have dimensions N1×N2. The vertical displacement W 1250 may be Figure 1 The vertical displacement determined in step 140 of method 100 or by combining the above Figure 7 , Figure 8 The vertical displacement determined by the process described. The horizontal displacement U, V 1260 can be Figure 1 The horizontal displacement determined in step 130 of method 100 or by combining the above Figure 2-Figure 6 The horizontal displacements (e.g., the first horizontal displacement and the second horizontal displacement) determined by the described process. In some embodiments, the vertical displacement W 1250 and the horizontal displacements U, V 1260 may be mapped to N1×N2 directions (i.e., the vertical dimension and the horizontal dimension are mapped to N1×N2 dimensions) and normalized using a Softmax function to obtain an N1×N2 dimensional vector, which may be used to weight the directional convolution result 1240.

[0158] Subsequently, at step 1280, the directional convolution result 1240 may be weighted by the weight 1275 to obtain a first modified deep learning feature 1285. The first modified deep learning feature 1285 may have dimensions [T, Z, C, X, Y]. The first modified deep learning feature 1285 may be subjected to feature transformation to obtain a predicted deep learning feature. The predicted deep learning feature may be Figure 1 The predicted deep learning features in step 180 of method 100 can be used to perform weather forecasting.

[0159] The following will combine Fig.13 Implementation details of the eighth process 1200 are further described.

[0160] Fig.13 1 is a schematic diagram of a ninth process 1300 for performing directional convolution operations on deep learning features and weighting directional convolution results according to some embodiments of the present invention. The ninth process 1300 may provide Figure 1 Implementation details of steps 150 and 160 of method 100, Fig. 9 The fifth process 900 is combined Fig.10 Implementation details of the sixth process 1000 in Fig. 9 The fifth process 900 is combined Fig.11 Implementation details of the seventh process 1100 or Fig.12 Implementation details of the eighth process 1200 in.

[0161] First, a second deep learning feature 1310 is provided. The second deep learning feature 1310 may be Figure 1 The second deep learning feature 1310 may be a time dimension, a vertical dimension, a channel dimension, and a horizontal dimension.

[0162] In the plane decomposition stage 1320, the feature space of the second deep learning feature 1310 can be decomposed into plane [Z, Y] and plane [Z, X]. Wherein, plane [Z, Y] can be defined by a vertical dimension and a horizontal dimension, and plane [Z, X] can be defined by a vertical dimension and another horizontal dimension. The second deep learning feature 1310 is correspondingly decomposed into a first feature 1332 and a second feature 1334. The plane decomposition stage 1320 can be Fig. 9 Step 910 of the fifth process 900 or Fig.12 Corresponding to step 1220 of the eighth process 1200.

[0163] Subsequently, in the directional convolution stage 1340, a directional convolution operation may be performed on the first feature 1332 using N1 directional convolution kernels 1351 on a plane to obtain a first directional convolution result 1353, where N1 is a positive integer. As an example, the first directional convolution result 1353 may have dimensions [T, X, N1×C, Z, Y]. Fig.14 The details of the convolution kernel used for the directional convolution operation are further described. The directional convolution operation can be performed on the second feature 1334 using N2 directional convolution kernels 1355 on another plane to obtain a second directional convolution result 1357, where N2 is a positive integer. As an example, the second directional convolution result 1357 can have dimensions [T, Y, N2×C, Z, X]. The directional convolution stage 1340 can be used with Fig. 9 Step 930, step 950 or step 960 of the fifth process 900 Fig.12 The steps 1226 and 1228 of the eighth process 1200 correspond to each other.

[0164] Subsequently, in the plane combination stage 1360, the first direction convolution result 1353 and the second direction convolution result 1357 may be combined to obtain a third direction convolution result 1370. As an example, the third direction convolution result 1370 may have dimensions [T, Z, N1×N2, C, X, Y]. The plane combination stage 1360 may be combined with Fig. 9 Step 970 of the fifth process 900 or Fig.12 Corresponding to step 1230 of the eighth process 1200.

[0165] Subsequently, in a weighting stage 1380, a weight 1395 may be determined based on the vertical displacement W 1391. The vertical displacement W 1391 may be Figure 1 The vertical displacement determined at step 140 of method 100 or by the above Figure 7 , Figure 8 Alternatively, in some embodiments, the weight 1395 may be determined based on the vertical displacement W 1391 and the horizontal displacements U, V 1393. The horizontal displacements U, V 1393 may be Figure 1 The horizontal displacement determined at step 130 of method 100 or by Figure 2-Figure 6 The described process determines the horizontal displacement. As an example, weights 1395 may have dimensions N1×N2.

[0166] In some embodiments, the vertical displacement W 1391 can be mapped to N1×N2 directions (i.e., the vertical dimension is mapped to N1×N2 dimensions) and normalized using a Softmax function to obtain N1×N2 vectors, which can be used to weight the third direction convolution result 1370.

[0167] In some embodiments, the vertical displacement W 1391 and the horizontal displacements U, V 1393 may be mapped to N1×N2 directions (i.e., the vertical dimension and the horizontal dimension are mapped to N1×N2 dimensions) and normalized using a Softmax function to obtain an N1×N2-dimensional vector, which may be used to weight the directional convolution result 1395.

[0168] Subsequently, the third directional convolution result 1370 may be weighted by the weight 1395 to obtain a second modified deep learning feature 1397. The second modified deep learning feature 1397 may be used for weather forecasting. As an example, the second modified deep learning feature 1397 may have dimensions [T, Z, C, X, Y]. The second modified deep learning feature 1397 may be subjected to feature transformation to obtain a predicted deep learning feature. The predicted deep learning feature may be Figure 1The predicted deep learning features in step 180 of method 100 can be used to make weather forecasts.

[0169] The weighting stage 1380 can be used with Fig.10 Sixth course in 1000 or Fig.11 Seventh course 1100 or Fig.12 Corresponding to step 1280 in the eighth process 1200.

[0170] Fig.14 is a schematic diagram of a convolution kernel used for directional convolution operation according to some embodiments of the present invention. Fig.14 The various convolution kernels shown in Figure 1 The directional convolution operation in step 150 of the method 100, Fig. 9 Step 930 of the fifth process 900, the directional convolution operation in step 950, Fig.12 Step 1226 of the eighth process 1200, the directional convolution operation in step 1228, Fig.13 The directional convolution operation in the directional convolution stage 1340 of the ninth process 1300 is described above, but the scope of the present invention is not limited to this.

[0171] Directional convolution operation refers to the convolution operation performed using a convolution kernel in a one-dimensional direction on a two-dimensional plane. Directional convolution operation can be used to extract the features of the input features in a specific direction.

[0172] The convolution kernel in the longitudinal direction is shown at (a), which is centered on the cell 1410 located in the center and extends to the upper and lower edges in both directions along the longitudinal direction.

[0173] (b) shows the convolution kernel in the lateral direction. (c) and (d) show the convolution kernel in the diagonal direction.

[0174] At (e), the convolution kernel is shown extending from the left to the right in the horizontal direction as a whole, and at (f), the convolution kernel is shown extending from the top to the bottom in the vertical direction as a whole.

[0175] At (g) the convolution kernel is shown extending from the upper side to the right after being turned.

[0176] Another convolution kernel is shown at (h) extending entirely from the left to the right in the horizontal direction.

[0177] The various convolution kernels shown are only exemplary. Embodiments may also use any other convolution kernels to perform directional convolution operations.

[0178] According to another aspect of the present invention, a computer device for weather forecasting is provided.

[0179] Fig.15FIG. 1 is a block diagram of a computer device 1500 for weather forecasting according to some embodiments of the present invention. The computer device 1500 may be used to implement the above-mentioned Figure 1-Figure 14 Describe the methods and procedures.

[0180] The computer device 1500 may include a memory 1510 and a processor 1520. The memory 1510 stores a computer program 1515. The processor 1520 may execute the computer program 1515 to implement the above-mentioned Figure 1-Figure 14 Describe the methods and procedures.

[0181] According to another aspect of the present invention, a system for weather forecasting is provided.

[0182] Fig.16 A block diagram of a system 1600 for weather forecasting according to some embodiments of the present invention is shown.

[0183] System 1600 may include computer device 1500 and detection device 1620. Computer device 1500 may be Fig.15 The computer device 1500 in FIG.

[0184] The detection device 1620 may be coupled to the computer device 1500. The detection device 1620 may be used to detect meteorological data. In some embodiments, the detection device 1620 may include a weather radar, which may be used to detect radar echo images, including reflectivity data. In some embodiments, the detection device 1620 may include various meteorological sensors, which may be used to detect various atmospheric variables, such as atmospheric pressure, atmospheric temperature, etc.

[0185] According to another aspect of the present invention, a computer-readable storage medium is provided.

[0186] Fig.17 A block diagram of a computer-readable storage medium 1700 is shown according to some embodiments of the invention.

[0187] The computer readable storage medium 1700 may store a computer program 1515. When the computer program 1515 is executed by the processor, the computer program 1515 may implement the above-mentioned Figure 1-Figure 14 Describe the methods and procedures.

[0188] According to another aspect of the present invention, a computer program product is provided.

[0189] Fig.18 A block diagram of a computer program product 1800 is shown according to some embodiments of the invention.

[0190] The computer readable storage medium 1800 may include a computer program 1515. When the computer program 1515 is executed by a processor, the computer program 1515 may implement the above-mentioned Figure 1-Figure 14Describe the methods and procedures.

[0191] Embodiments of the present invention have been described with reference to the accompanying drawings. The embodiments are illustrative rather than restrictive.

Claims

1. A method for weather forecasting, characterized in that: The method comprises: Obtain weather data; Extracting deep learning features from the meteorological data, the deep learning features comprising a plurality of frames in a feature space, the feature space comprising a time dimension and a spatial dimension, the spatial dimension comprising a horizontal dimension within a plane at a predetermined height relative to a reference plane and a vertical dimension perpendicular to the reference plane; estimating, in the feature space, a horizontal displacement corresponding to a movement of a reference unit in a first frame of the plurality of frames from a reference position in the first frame to a target position in a second frame of the plurality of frames, wherein the second frame is adjacent to the first frame in the time dimension, and the horizontal displacement is associated with the horizontal dimension; Determining a vertical displacement of the reference unit moving from the reference position to the target position according to a three-dimensional continuity equation in the feature space based on the horizontal displacement, the vertical displacement being associated with the vertical dimension; Performing a directional convolution operation on the deep learning feature using multiple directional convolution kernels on a plane to determine a directional convolution result; The directional convolution result is weighted based on the vertical displacement to determine a modified deep learning feature, and the modified deep learning feature is used to predict meteorological data at a future time.

2. The method according to claim 1, characterized in that The meteorological data is obtained by selecting raw meteorological data at preset height intervals between a preset initial height and an end height.

3. The method according to claim 2, characterized in that The raw meteorological data include radar echo images or atmospheric variables obtained at different altitudes.

4. The method according to claim 1, characterized in that Extracting deep learning features includes: performing downsampling on the meteorological data.

5. The method according to claim 1, wherein The horizontal dimension includes a first horizontal dimension and a second horizontal dimension, and Estimated horizontal displacement includes: estimating a first horizontal displacement, the first horizontal displacement associated with the first horizontal dimension; and A second horizontal displacement is estimated, the second horizontal displacement being associated with the second horizontal dimension.

6. The method according to claim 1, characterized in that Estimated horizontal displacement includes: calculating feature similarity of the reference cell relative to cells at each potential target location in the neighborhood in the second frame; weighting a potential horizontal displacement of the reference unit from the reference position to each potential target position in the neighborhood using the feature similarity; and The horizontal displacement is estimated as a weighted sum of the potential horizontal displacements associated with each potential target position.

7. The method according to claim 1, characterized in that Estimated horizontal displacement includes: performing optical flow deformation on the first frame after the reference unit is moved by a potential horizontal displacement; and The horizontal displacement is estimated by minimizing the error between the second frame and the first frame deformed by optical flow.

8. The method according to claim 1, characterized in that Determining the vertical displacement of the reference unit from the reference position to the target position includes: Determining, based on the horizontal displacement, a vertical partial differential component of the reference unit moving from the reference position to the target position according to a three-dimensional continuity equation in the feature space; and The vertical partial differential component is integrated in the vertical dimension to determine the vertical displacement.

9. The method according to claim 1, wherein: The horizontal dimension includes a first horizontal dimension and a second horizontal dimension, and Performing a directional convolution operation on the deep learning feature includes: decomposing the spatial dimension into a first plane defined by the first horizontal dimension and the vertical dimension and a second plane defined by the second horizontal dimension and the vertical dimension; Performing directional convolution operations on the deep learning features in first multiple directions on the first plane to obtain first multiple directional convolution results; Performing directional convolution operations on the deep learning features in second multiple directions on the second plane to obtain second multiple directional convolution results; and The first plurality of directional convolution results and the second plurality of directional convolution results are combined to determine the directional convolution result.

10. The method according to claim 9, characterized in that Weighting the directional convolution result includes: determining a first plurality of weights and a second plurality of weights based on the vertical displacement, the first plurality of weights being associated with the first plurality of directions and the second plurality of weights being associated with the second plurality of directions; and The directional convolution result is weighted using the first plurality of weights and the second plurality of weights.

11. The method according to claim 9, characterized in that Weighting the directional convolution result includes: determining a first plurality of weights and a second plurality of weights based on the vertical displacement and the horizontal displacement, the first plurality of weights being associated with the first plurality of directions and the second plurality of weights being associated with the second plurality of directions; and The directional convolution result is weighted using the first plurality of weights and the second plurality of weights.

12. The method according to claim 1, wherein: The method further comprises: performing feature extraction on the modified deep learning features to determine predicted deep learning features; and Perform weather forecast based on the predicted deep learning features.

13. A computer device for weather forecasting, comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the method according to any one of claims 1-12.

14. A system for weather forecasting, characterized in that: The system comprises: The computer device of claim 13; and The detection device is coupled to the computer device and is used to detect the meteorological data.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 12 are implemented.

16. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 12 are implemented.

Images