Non-linear short-term and imminent forecasting method, device and equipment based on radar data

Abstract

The invention provides a nonlinear short-term and imminent forecasting method, device and equipment based on radar data, and the method comprises the steps: obtaining historical radar echo data before a to-be-predicted moment, and carrying out the preprocessing of the radar echo data, and obtaining a radar echo map; substituting the radar echo map into a pre-trained dual-channel convolution time sequence extrapolation model to obtain radar extrapolation prediction data; wherein the pre-trained dual-channel convolution time sequence extrapolation model is obtained by adopting deep learning convolution to learn a radar echo map before a historical moment and a radar echo map after the historical moment; the dual-channel convolution time sequence extrapolation model comprises a deep learning convolution dual-channel encoder and a decoder. By using the radar extrapolation method, forecasting and early warning of short-time severe convection weather can be efficiently and accurately realized, so that the loss of a transformer substation and a power transmission line caused by local severe convection meteorological disasters is reduced, and the emergency response capability of power equipment and facilities is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and data-driven technology, and in particular to a nonlinear short-term forecasting method, device and equipment based on radar data. Background Art

[0002] From the perspective of meteorology, short-term forecasts usually refer to short-term weather forecasts within 12 hours and near-term weather forecasts within 2 hours. In terms of forecast timeliness, they are the shortest weather forecasts. The characteristics of short-term severe convective weather include short-term, high intensity, and suddenness. It appears suddenly in local areas, develops rapidly, and then disappears quickly. It can cause power system equipment to trip, power outages, tower collapse, and even cause casualties and other accidents, causing huge economic and social losses. The current mesoscale numerical forecast model can have good prediction results in short-term and medium-term weather forecasts, while the short-term severe convective process belongs to the small and micro-scale weather system. The numerical weather forecast model has a large amount of calculation, and consumes a lot of computing resources and time. Its ability to respond to sudden weather is weak. The observation method for severe convective weather mainly relies on weather radar. Through weather radar data, meteorological staff can observe the system structure and movement field of convection, thereby realizing the identification of thunderstorm weather.

[0003] The traditional linear extrapolation method based on radar data has a short effective forecast period, and it is difficult to exceed 30 minutes for a high-accuracy forecast period. In addition, the formation and development process of thunderstorm clouds is generally not taken into account during the radar extrapolation process, resulting in serious false alarms. Summary of the invention

[0004] In order to solve the problem that the linear extrapolation method based on radar data in the prior art has a short effective time of prediction and there is a serious false alarm phenomenon in the radar extrapolation process, the present invention proposes a nonlinear short-term prediction method based on radar data, comprising:

[0005] Acquire historical radar echo data before the time to be predicted, and preprocess the radar echo data to obtain a radar echo map;

[0006] Substituting the radar echo map into a pre-trained dual-channel convolutional time series extrapolation model to obtain radar extrapolation prediction data;

[0007] The pre-trained dual-channel convolutional time series extrapolation model is obtained by learning the radar echo map before a certain moment in history and the radar echo map after the certain moment by using deep learning convolution;

[0008] The dual-channel convolutional timing extrapolation model includes a deep learning convolutional dual-channel encoder and decoder.

[0009] Optionally, the acquiring of historical radar echo data before the time to be predicted and preprocessing the radar echo data to obtain a radar echo map includes:

[0010] Acquire radar echo data within a preset period before the time to be predicted;

[0011] Extracting information from the radar echo data to obtain the latitude and longitude range of the area covered by the radar echo;

[0012] The longitude and latitude range of the coverage area of ​​the radar echo is clipped according to the set spatial resolution to obtain a clipped radar echo map;

[0013] The data in the clipped radar echo map is quality screened to obtain a high-quality radar echo map.

[0014] Optionally, extracting information from the radar echo data to obtain the latitude and longitude range of the coverage area of ​​the radar echo includes:

[0015] Calculate the longitude and latitude coordinates of each grid point from the data area description parameters in the file header of the radar echo data;

[0016] Based on the description of the radar file in the file header, obtain the latitude and longitude range of the area covered by the radar echo.

[0017] Optionally, the longitude and latitude range of the coverage area of ​​the radar echo is clipped according to a set spatial resolution to obtain a clipped radar echo map, including:

[0018] Determine the core area of ​​current nowcast modeling based on the latitude and longitude range of the coverage area of ​​the radar echo;

[0019] The radar echo map is cropped with the central longitude and latitude of the core area as the center and with the set spatial resolution to obtain a cropped radar echo map.

[0020] Optionally, the performing quality screening on the data in the clipped radar echo map to obtain a high-quality radar echo map includes:

[0021] The points whose radar echo values ​​in the clipped radar echo map satisfy the first relational expression are regarded as isolated high values;

[0022] The points in the clipped radar echo map whose radar echo values ​​satisfy the second relational expression are regarded as ground object echo points;

[0023] Isolated high values ​​and ground object echo points are removed from the clipped radar echo map to obtain a high-quality radar echo map.

[0024] Optionally, the first relationship is as follows:

[0025]

[0026] Where M 5*5 is a 5×5 grid range within the surrounding area, M i The 5×5 grid range M within the surrounding area 5 *5 The radar echo value of any point except the center point within x,y is a point in the radar echo graph.

[0027] Optionally, the second relationship is as follows:

[0028]

[0029] Where N 3*3 is a 3×3 grid range within the surrounding area, N i The 3×3 grid range N within the surrounding area 3 *3 The radar echo value of any point within.

[0030] Optionally, the training of the dual-channel convolutional timing extrapolation model includes:

[0031] Obtain the historical radar echo map of the area to be measured;

[0032] Constructing a sample set from the historical radar echo graphs before a certain moment and the historical radar echo graphs after the certain moment;

[0033] Dividing the sample set into a training set and a test set according to a set ratio;

[0034] The historical radar echo map before the certain moment is used as input, and the historical radar echo map after the certain moment is used as output to train the dual-channel convolution time series extrapolation model, so as to obtain a preliminarily trained dual-channel convolution time series extrapolation model;

[0035] The preliminarily trained dual-channel convolution timing extrapolation model is tested based on the test set, and the preliminarily trained dual-channel convolution timing extrapolation model that passes the test is used as the trained dual-channel convolution timing extrapolation model.

[0036] Optionally, the testing the preliminarily trained dual-channel convolution timing extrapolation model based on the test set, and using the preliminarily trained dual-channel convolution timing extrapolation model that passes the test as the trained dual-channel convolution timing extrapolation model, includes:

[0037] Substituting the historical radar echo map before a certain moment in the test set into the preliminarily trained dual-channel convolutional time series extrapolation model to obtain the radar extrapolation prediction data at the said certain moment;

[0038] Calculate the evaluation index value based on the radar extrapolation prediction data at the certain moment combined with the historical radar data after a certain moment in the test set;

[0039] Based on whether the evaluation index value meets a preset threshold, if so, the test is passed, and the preliminarily trained dual-channel convolution timing extrapolation model is used as the trained dual-channel convolution timing extrapolation model; otherwise, the test is not passed;

[0040] The evaluation index values ​​include: correlation coefficient, root mean square error, and image structure similarity index.

[0041] Optionally, the construction of the dual-channel convolutional time series extrapolation model includes:

[0042] An encoder is constructed by a time sub-network for extracting time data, a space sub-network for extracting space data, and a spatiotemporal data fusion module for fusing the time data with the space data, wherein the time sub-network and the space sub-network are in parallel;

[0043] A dual-channel convolutional temporal extrapolation model is constructed by an encoder and a decoder that decodes high-bit fused features into predicted future spatiotemporal data.

[0044] Optionally, substituting the radar echo map into a pre-trained dual-channel convolution time series extrapolation model to obtain radar extrapolation prediction data includes:

[0045] extracting time features and space features from the processed radar echo data through a time sub-network and a space sub-network;

[0046] The temporal feature and the spatial feature are fused by a spatiotemporal data fusion module to obtain a fused feature;

[0047] The fused features are decoded by a decoder, and the fused features are converted into radar extrapolation prediction data.

[0048] On the other hand, the present invention also provides a nonlinear short-term forecasting device based on radar data, comprising:

[0049] A parameter processing module, used for acquiring historical radar echo data before the time to be predicted, and preprocessing the radar echo data to obtain a radar echo map;

[0050] A prediction module, used for substituting the radar echogram into a pre-trained dual-channel convolution time series extrapolation model to obtain radar extrapolation prediction data;

[0051] The pre-trained dual-channel convolutional time series extrapolation model is obtained by learning the radar echo map before a certain moment in history and the radar echo map after the certain moment by using deep learning convolution;

[0052] The dual-channel convolutional timing extrapolation model includes a deep learning convolutional dual-channel encoder and decoder.

[0053] Optionally, the parameter processing module includes:

[0054] The acquisition submodule is used to acquire radar echo data within a preset period before the time to be predicted;

[0055] An extraction submodule, used to extract information from the radar echo data to obtain the latitude and longitude range of the coverage area of ​​the radar echo;

[0056] A clipping submodule, used to clip the latitude and longitude range of the coverage area of ​​the radar echo according to a set spatial resolution to obtain a clipped radar echo map;

[0057] The screening submodule is used to perform quality screening on the data in the clipped radar echo map to obtain a high-quality radar echo map.

[0058] Optionally, the extraction submodule is specifically used for:

[0059] Calculate the longitude and latitude coordinates of each grid point from the data area description parameters in the file header of the radar echo data;

[0060] Based on the description of the radar file in the file header, obtain the latitude and longitude range of the area covered by the radar echo.

[0061] Optionally, the cutting submodule is specifically used for:

[0062] Determine the core area of ​​current nowcast modeling based on the latitude and longitude range of the coverage area of ​​the radar echo;

[0063] The radar echo map is cropped with the central longitude and latitude of the core area as the center and with the set spatial resolution to obtain a cropped radar echo map.

[0064] Optionally, the screening submodule is specifically used for:

[0065] The points whose radar echo values ​​in the clipped radar echo map satisfy the first relational expression are regarded as isolated high values;

[0066] The points in the clipped radar echo map whose radar echo values ​​satisfy the second relational expression are regarded as ground object echo points;

[0067] Isolated high values ​​and ground object echo points are removed from the clipped radar echo map to obtain a high-quality radar echo map.

[0068] Optionally, the first relationship is as follows:

[0069]

[0070] Where M 5*5 is a 5×5 grid range within the surrounding area, M i The 5×5 grid range M within the surrounding area 5 *5 The radar echo value of any point except the center point within x,y is a point in the radar echo graph.

[0071] Optionally, the second relationship is as follows:

[0072]

[0073] Where N 3*3 is a 3×3 grid range within the surrounding area, N i The 3×3 grid range N within the surrounding area 3 *3 The radar echo value of any point within.

[0074] Optionally, a training module is also included, for:

[0075] Obtain the historical radar echo map of the area to be measured;

[0076] Constructing a sample set from the historical radar echo graphs before a certain moment and the historical radar echo graphs after the certain moment;

[0077] Dividing the sample set into a training set and a test set according to a set ratio;

[0078] The historical radar echo map before the certain moment is used as input, and the historical radar echo map after the certain moment is used as output to train the dual-channel convolution time series extrapolation model, so as to obtain a preliminarily trained dual-channel convolution time series extrapolation model;

[0079] The preliminarily trained dual-channel convolution timing extrapolation model is tested based on the test set, and the preliminarily trained dual-channel convolution timing extrapolation model that passes the test is used as the trained dual-channel convolution timing extrapolation model.

[0080] Optionally, building blocks are also included for:

[0081] An encoder is constructed by a time sub-network for extracting time data, a space sub-network for extracting space data, and a spatiotemporal data fusion module for fusing the time data with the space data, wherein the time sub-network and the space sub-network are in parallel;

[0082] A dual-channel convolutional temporal extrapolation model is constructed by an encoder and a decoder that decodes high-bit fused features into predicted future spatiotemporal data.

[0083] In another aspect, the present application further provides an electronic device, comprising: at least one processor and a memory; the memory and the processor are connected via a bus;

[0084] A memory for storing one or more programs;

[0085] When one or more programs are executed by at least one processor, the above-mentioned nonlinear short-term prediction method based on radar data is implemented.

[0086] On the other hand, the present application further provides a readable storage medium having an execution program stored thereon, and when the execution program is executed, the nonlinear short-term and impending prediction method based on radar data as described above is implemented.

[0087] Compared with the prior art, the present invention has the following beneficial effects:

[0088] The present invention provides a nonlinear short-term forecasting method, device and equipment based on radar data, including: obtaining historical radar echo data before the time to be predicted, and preprocessing the radar echo data to obtain a radar echo map; substituting the radar echo map into a pre-trained dual-channel convolution time series extrapolation model to obtain radar extrapolation prediction data; wherein the pre-trained dual-channel convolution time series extrapolation model is obtained by using deep learning convolution to learn the radar echo map before a certain historical moment and the radar echo map after the certain moment; the dual-channel convolution time series extrapolation model includes a deep learning convolution dual-channel encoder and decoder. The present invention utilizes the radar extrapolation method to efficiently and accurately realize the forecast and warning of short-term severe convective weather, thereby reducing the losses caused by local severe convective meteorological disasters to substations and transmission lines, and improving the emergency response capabilities of power equipment and facilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 It is a flow chart of the nonlinear short-term forecasting method based on radar data of the present invention;

[0090] Figure 2 Schematic diagram of the structure of the backbone convolution block and the residual convolution block of the present invention;

[0091] Figure 3 A radar echo time series feature graph extracted by the time sub-network of the present invention;

[0092] Figure 4 It is a spatial structure feature graph extracted based on the spatial sub-network of the present invention;

[0093] Figure 5 It is a flowchart of the processing of the backbone convolution block and the residual convolution block of the present invention;

[0094] Figure 6 It is a schematic diagram of the structure of the spatiotemporal data fusion module of the present invention;

[0095] Figure 7 A comparison chart of the radar extrapolation image prediction and the actual in Jiangsu and Shanghai areas in the embodiment of the present invention;

[0096] Figure 8 The figure is a schematic diagram of the structure of an electronic device of the present invention. DETAILED DESCRIPTION

[0097] Based on radar data, radar feature extraction is performed, and deep learning image processing technology and encoding and decoding model construction technology are used to achieve accurate short-term forecasts of 0 to 6 hours for short-term severe convective weather. Through a nonlinear short-term forecast method based on radar data proposed in this application, it is possible to achieve short-term rainfall forecasts based on a data-driven method, explore the short-term variation laws of historical radar echo data, and achieve high timeliness and high accuracy forecasts of short-term severe convective weather.

[0098] In order to better understand the present invention, the content of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0099] Example 1

[0100] The present invention proposes a nonlinear short-term forecasting method based on radar data, such as Figure 1 As shown, including:

[0101] Step I: acquiring historical radar echo data before the time to be predicted, and preprocessing the radar echo data to obtain a radar echo map;

[0102] Step II: Substituting the radar echo map into a pre-trained dual-channel convolution time series extrapolation model to obtain radar extrapolation prediction data;

[0103] The pre-trained dual-channel convolutional time series extrapolation model is obtained by learning the radar echo map before a certain moment in history and the radar echo map after the certain moment by using deep learning convolution;

[0104] The dual-channel convolutional timing extrapolation model includes a deep learning convolutional dual-channel encoder and decoder.

[0105] The present invention provides a radar extrapolation short-term forecasting method based on a dual-channel 3D convolutional neural network architecture, the main steps of which include radar data analysis, radar data processing, model construction, and model training. The following is a further introduction to each step of the present invention:

[0106] Step I: Acquire historical radar echo data before the time to be predicted, and preprocess the radar echo data to obtain a radar echo map, including:

[0107] Acquire radar echo data within a preset period before the time to be predicted;

[0108] Extracting information from the radar echo data to obtain the latitude and longitude range of the area covered by the radar echo;

[0109] The longitude and latitude range of the coverage area of ​​the radar echo is clipped according to the set spatial resolution to obtain a clipped radar echo map;

[0110] The data in the clipped radar echo map is quality screened to obtain a high-quality radar echo map.

[0111] Furthermore, the information extraction of the radar echo data to obtain the latitude and longitude range of the coverage area of ​​the radar echo includes:

[0112] Calculate the longitude and latitude coordinates of each grid point from the data area description parameters in the file header of the radar echo data;

[0113] Based on the description of the radar file in the file header, obtain the latitude and longitude range of the area covered by the radar echo.

[0114] Furthermore, the longitude and latitude range of the coverage area of ​​the radar echo is clipped according to the set spatial resolution to obtain a clipped radar echo map, including:

[0115] Determine the core area of ​​current nowcast modeling based on the latitude and longitude range of the coverage area of ​​the radar echo;

[0116] The radar echo map is cropped with the central longitude and latitude of the core area as the center and with the set spatial resolution to obtain a cropped radar echo map.

[0117] Furthermore, the quality screening of the data in the clipped radar echo map to obtain a high-quality radar echo map includes:

[0118] The points whose radar echo values ​​in the clipped radar echo map satisfy the first relational expression are regarded as isolated high values;

[0119] The points in the clipped radar echo map whose radar echo values ​​satisfy the second relational expression are regarded as ground object echo points;

[0120] Isolated high values ​​and ground object echo points are removed from the clipped radar echo map to obtain a high-quality radar echo map.

[0121] Furthermore, the first relationship is as follows:

[0122]

[0123] Where M 5*5 is a 5×5 grid range within the surrounding area, M i The 5×5 grid range M within the surrounding area 5 *5 The radar echo value of any point except the center point within x,y is a point in the radar echo graph.

[0124] Furthermore, the second relationship is as follows:

[0125]

[0126] Where N 3*3 is a 3×3 grid range within the surrounding area, N i The 3×3 grid range N within the surrounding area 3 *3 The radar echo value of any point within.

[0127] Step I specifically includes:

[0128] (1) Radar data extraction.

[0129] The original radar echo data is mostly binary data files. The equal longitude and latitude grid data saved by radar data consists of a file header and a data area. The file header structure definition refers to the table below. Data storage method: The jigsaw puzzle data of the equal longitude and latitude grid is enlarged, shortened, and compressed for storage. Data arrangement method: Arranged row by row from north to south, and distributed column by column from west to east. The longitude and latitude coordinates of each grid point can be calculated based on the parameters describing the data area of ​​the file header. Based on the description of the radar file in the file header, the longitude and latitude range of the coverage area of ​​the radar echo can be obtained, and the real radar echo value can be obtained based on the magnification.

[0130] (2) Radar Data Center Cutting

[0131] Determine the core area of ​​the current nowcast modeling. For all radar echo data, crop the center of the current area with the central longitude and latitude as the center, and construct a 256*256 grid with a spatial resolution of 0.1 degrees longitude and latitude. Use this grid as the input value of the nowcast model.

[0132] (3) Radar data processing

[0133] The radar data quality is controlled. There are isolated high-value noise and interference from ground objects in the radar echo. The data processing link is to remove the isolated high-value and ground clutter of the radar echo. For a certain point, the echo value is P x,y , Px,y The 5×5 square around the area excluding P x,y The echo value set is recorded as M 5*5 = {P x-i,y-j},and Where i,j∈{-1,-2,0,1,2}. x,y The echo value in the surrounding 3×3 square is recorded as N 3*3 = {P x-i,y-j}, where i,j∈{-1,0,1}.

[0134]

[0135] 1. For a point P in the radar echo graph x,y , the 5×5 grid range M within its surrounding area 5*5 Except the center point P x,y The radar echo value M at any point i , meeting the requirement of (1), the echo at this point is an isolated high value.

[0136] 2. For a point P in the radar echo graph x,y , the 3×3 grid range N within its surrounding area 3*3 The radar echo value N of any point within i , which meets the requirement of (2), and this point belongs to the ground object echo point.

[0137] Any item that meets 1 or 2 above is an interference item of radar data that can be removed. Abnormal data and invalid data of radar echo data are eliminated based on rules to improve the data quality of radar echo data.

[0138] Before step II, the method also includes: constructing a dual-channel convolution timing extrapolation model, training the dual-channel convolution timing extrapolation model, and obtaining a trained dual-channel convolution timing extrapolation model.

[0139] (1) Model construction

[0140] The model structure described in this method is a dual-channel convolutional temporal extrapolation model. The basic unit of the network is 3DCNN. The main steps of radar extrapolation implemented by it are to input the radar of historical T frames, encode the historical radar data using a deep learning convolutional dual-channel encoder, extract high-dimensional features, and then restore it to the original image size based on the decoder to predict the result of radar extrapolation.

[0141] The overall structure of the model is as follows Figure 2 As shown, the encoder has two parallel paths for spatiotemporal feature extraction with intermediate fusion, and the decoder is used to decode the high-order fused features into predicted future spatiotemporal data:

[0142] 1. Input data: Input the radar echo map of T frames before the current moment ∏ T =(I X-T+1 ,...,I X ), where I X Represents the radar echo map at the current time X. Each echo map is a 256*256 two-dimensional matrix after center cropping. T is the number of frames input in advance, which is a positive integer. T The sequence is formed by sampling every 4 frames S , that is, S = 1 / 4T, the input tensor structure of the time subnetwork is (B, T, H, W), and the input data structure of the spatial subnetwork is (B, S, H, W), where S represents the total number of sequences, B represents the number of samples input during training, and H and W represent the length and height of the input radar image.

[0143] 2. Time subnetwork: The time subnetwork is mainly used to capture the time intensity evolution characteristics of radar echoes, such as Figure 3 As shown in the figure, center radar intensity represents the center radar intensity, intensity / dbz represents the intensity, pred represents the predicted value, and real represents the actual value. The network structure is 1 trunk convolution block (Trunk) and 2 convolution residual blocks (Res). For the specific network structure, refer to Figure 5 The main convolutional blocks are composed of 1 convolutional layer, 1 BN layer, 1 activation function layer and a Max Pooling layer. The specific structure is shown in Figure 5 Each convolution residual block includes multiple residual modules, where the convolution change of the residual module is recorded as F(·). The mathematical relationship of a single residual module can be expressed as formula (3). The size of the echo image before and after the residual module does not change. The input of the time subnetwork is ∏ T , the time length of a single sample is T. The time sub-network adopts a smaller convolution kernel, which can have a finer receptive field and a faster update frequency. It mainly focuses on the intensity changes of specific points to realize the prediction of echo intensity.

[0144]

[0145] Where l>=2

[0146] In the formula, is the feature extracted from the first layer, is the feature extracted from the lth layer, Represents the extracted feature value of the l-1th layer, Trunk represents the trunk convolution block, and Res represents the residual convolution block.

[0147] 3. Spatial subnetwork: The spatial subnetwork has the same structure as the temporal subnetwork, but different specific parameters. It is mainly used to capture the spatial deformation and displacement characteristics of radar echoes, such as Figure 4 As shown in the figure, the network structure is 1 trunk convolution block (Trunk) and 3 convolution residual blocks (Res). The trunk convolution block consists of 1 convolution layer, 1 BN layer, 1 activation function layer and a Max Pooling layer. The input of the spatial subnetwork is ∏ S The time length of a single sample is S = 1 / 4T, and its input frequency is 1 / 4 of the spatial sub-network. It uses a larger convolution kernel, has a larger receptive field, and a lower input frequency. It mainly extracts the image and contour features of the radar echo.

[0148] 4. Input the radar echo map of the current time sub-network into the sample ∏ T The sampling sequence is formed by sampling every 4 frames. S , the sampling sequence ∏ S Input to the spatial sub-network, and capture the spatial and displacement change characteristics of the radar echo map through convolution calculation, S =(I X-T ,I X-T+4 ,...,I X ) The input time length of the radar echo map of the spatial subnetwork is S = 1 / 4T, and the spatial and displacement change characteristics It is expressed as:

[0149]

[0150] Where l>=2

[0151] In the formula, Trunk represents the trunk convolution block, Res represents the residual convolution block, Represents the spatial feature extraction result of the current layer, represents the spatial feature extraction result of the previous layer, x l-1 Represents the fusion result of the previous layer of spatiotemporal features.

[0152] The spatiotemporal data fusion module (fuse) is as follows Figure 6 As shown in the figure, C represents the fusion module, K is the convolution kernel, and K f Represents the convolution size of the time dimension: The spatial and temporal information of the two sub-networks are fused at each convolution downscaling step of the encoder. The fusion strategy is to superimpose the intensity change characteristics in time on the basis of capturing the spatial movement. Due to the different sizes of the convolution kernels of the spatiotemporal sub-networks, the sizes of the encoded temporal and spatial features are different. At this time, the data coupling module needs to process the sizes of the two to the same size. The spatial features extracted by the lth residual block are The feature obtained by time encoding is The mathematical expression of this module is shown in formula (4).

[0153]

[0154] in, It represents a 3D convolution with a convolution kernel of kf*1*1, followed by the sequential operation of padding, BN and Relu layers. kf is a selectable hyperparameter that converts temporal features into motion feature sizes. Concat represents a concatenation operation that realizes the fusion of spatiotemporal features by concatenating spatiotemporal features.

[0155] 5.Decoder:

[0156] The main function of the decoder is to convert the high-dimensional spatiotemporal features constructed by the encoder into the final prediction graph. The features are sequentially input into three spatial decoding modules, with the deconvolution module as the main means to achieve upsampling and expand the spatial resolution of the result. The formula of the decoding process is as follows:

[0157] y l ←Up_Pool(f tc (y l-1 ; W 1*kc*kc )(5)

[0158] Among them, y l represents the output of the lth spatial decoding module, y l-1 represents the output of the l-1th spatial decoding module, f tc (y l-1 ; W 1*kc*kc ) represents the sequential operation of 3D deconvolution, BN and ReLU, where the convolution kernel of deconvolution is 1*k c *k c , kc is determined according to the specific output and output size requirements. Up_Pool is the maximum up-pooling operation. The main function of deconvolution is to reduce the number of channels and expand the feature size, which is used to gradually restore the image size and realize radar extrapolation prediction.

[0159] R = gelu(f c (y3; W 2*1*1 ) (6)

[0160] The final output normalization result R is shown in formula (6), W 2*1*1 represents a 3D 1*1 convolution kernel, which is used to adjust the output size to make it consistent with the input size. R represents the final output result, and f c (y3; W 2*1*1) represents the sequential operation of 3D deconvolution, BN and ReLU.

[0161] 6. Loss function:

[0162] In order to better reflect the deviation between the predicted image and the actual image and improve the accuracy of fitting, three indicators, namely correlation coefficient, mean square error (MSE), and image structure similarity index (SSIM), are used to evaluate the prediction results:

[0163]

[0164]

[0165] Where X represents the predicted result, Y represents the true value, cov(X,Y) represents the covariance of X and Y, δ X represents the standard deviation of the X image matrix, δ Y represents the standard deviation of the Y image matrix, N is the number of pixels in the image, and u x 、u y Represent the means of the two images respectively, c1 and c2 are constants to avoid the denominator being 0, and are usually small.

[0166] Model training

[0167] Regarding the input and output settings of the network, we need to further process the data and set T=15, that is, 16 consecutive frames are used as the input of the model and 16 consecutive frames are used as the output. The time resolution is 30 minutes, and the radar extrapolation prediction of 6 hours is achieved.

[0168] Step II: Substitute the radar echo map into a pre-trained dual-channel convolutional time series extrapolation model to obtain radar extrapolation prediction data.

[0169] Example 2

[0170] This method is based on the radar echo data in Jiangsu Province as a case to verify the effectiveness and advancement of this method. The specific steps and implementation results are as follows:

[0171] Step 1: Count and integrate the radar echo data of Jiangsu and Shanghai in the past two years, and extract the radar echo data based on the above coding rules. Get the central longitude and latitude of Jiangsu and Shanghai (32.96N, 119.56E), and extract the radar echo data of the 256×256 area covering Jiangsu and Shanghai with a resolution of 0.01°.

[0172] Step 2: Based on the extracted radar echo data, remove isolated high-value noise and ground clutter according to the rules reflected in formulas (1) and (2).

[0173] Step 3: Based on Figure 2 The overall framework of the model shown and Figure 5 The specific structure of the model shown in Figure 1 constructs a dual-channel 3D convolutional radar extrapolation model. The training set is the radar echo data of the Jiangsu and Shanghai areas from 2021 to 2022 based on step 1. Take the first 15 frames at a certain moment as input, including the current frame, a total of 16 frames, the time channel takes T=15, the spatial channel S=(T+1) / 4-1=3, as the input of the model. And get the next 16 frames starting from the current moment as the prediction target of the model.

[0174] Step 4: Based on the training results of the model, judge the convergence of the model, and implement the extrapolation of radar echo data according to the trained model. The prediction effect of radar extrapolation prediction of severe convection is mainly evaluated by the following indicators. The test set is the heavy rain that occurred in Jiangsu Province in 2023. The specific case experimental results are as follows:

[0175] 1. Critical success index (CSI), the formula is:

[0176]

[0177] 2. Probability of Detection (POD), the formula is:

[0178]

[0179] 3. False Alarm Rate (FAR), the formula is:

[0180]

[0181] The meanings of TP, FN, and FP are as follows:

[0182] Table 1 Binary classification confusion matrix

[0183] Prediction is true Prediction is false Actual true True Position(TP) False Negative(FN) Actually false False Position(FP) True Negative(TN)

[0184] The prediction accuracy of the current model is shown in the table below, where CSI_25, POD_25, and FAR_25 only focus on points related to echoes greater than 25dbz. Similarly, indicators such as CSI_35 focus on the prediction accuracy of points greater than 35dbz. Step 1, step 8, and step 15 represent the prediction capabilities of the model in the early, mid-term, and late stages, respectively. The results show that this method is effective and has certain advantages in radar extrapolation problems, as shown in Table 2 and Figure 7 shown.

[0185] Table 2 Accuracy evaluation values

[0186] CSI_25 POD_25 FAR_25 step1 0.8663 0.9263 0.0865 step8 0.6985 0.8083 0.1629 step15 0.3305 0.4641 0.4656 CSI_35 POD_35 FAR_35 step1 0.7113 0.8409 0.1782 step8 0.4322 0.5688 0.3571 step15 0.3305 0.4641 0.4656

[0187] Example 3

[0188] The present invention also provides a nonlinear short-term forecasting device based on radar data, comprising:

[0189] A parameter processing module, used for acquiring historical radar echo data before the time to be predicted, and preprocessing the radar echo data to obtain a radar echo map;

[0190] A prediction module, used for substituting the radar echogram into a pre-trained dual-channel convolution time series extrapolation model to obtain radar extrapolation prediction data;

[0191] The pre-trained dual-channel convolutional time series extrapolation model is obtained by learning the radar echo map before a certain moment in history and the radar echo map after the certain moment by using deep learning convolution;

[0192] The dual-channel convolutional timing extrapolation model includes a deep learning convolutional dual-channel encoder and decoder.

[0193] Optionally, the parameter processing module includes:

[0194] The acquisition submodule is used to acquire radar echo data within a preset period before the time to be predicted;

[0195] An extraction submodule, used to extract information from the radar echo data to obtain the latitude and longitude range of the coverage area of ​​the radar echo;

[0196] A clipping submodule, used to clip the latitude and longitude range of the coverage area of ​​the radar echo according to a set spatial resolution to obtain a clipped radar echo map;

[0197] The screening submodule is used to perform quality screening on the data in the clipped radar echo map to obtain a high-quality radar echo map.

[0198] Optionally, the extraction submodule is specifically used for:

[0199] Calculate the longitude and latitude coordinates of each grid point from the data area description parameters in the file header of the radar echo data;

[0200] Based on the description of the radar file in the file header, obtain the latitude and longitude range of the area covered by the radar echo.

[0201] Optionally, the cutting submodule is specifically used for:

[0202] Determine the core area of ​​current nowcast modeling based on the latitude and longitude range of the coverage area of ​​the radar echo;

[0203] The radar echo map is cropped with the central longitude and latitude of the core area as the center and with the set spatial resolution to obtain a cropped radar echo map.

[0204] Optionally, the screening submodule is specifically used for:

[0205] The points whose radar echo values ​​in the clipped radar echo map satisfy the first relational expression are regarded as isolated high values;

[0206] The points in the clipped radar echo map whose radar echo values ​​satisfy the second relational expression are regarded as ground object echo points;

[0207] Isolated high values ​​and ground object echo points are removed from the clipped radar echo map to obtain a high-quality radar echo map.

[0208] Optionally, the first relationship is as follows:

[0209]

[0210] Where M 5*5 is a 5×5 grid range within the surrounding area, M i The 5×5 grid range M within the surrounding area 5 *5 The radar echo value of any point except the center point within x,y is a point in the radar echo graph.

[0211] Optionally, the second relationship is as follows:

[0212]

[0213] Where N 3*3 is a 3×3 grid range within the surrounding area, N i The 3×3 grid range N within the surrounding area 3 *3 The radar echo value of any point within.

[0214] Optionally, a training module is also included, for:

[0215] Obtain the historical radar echo map of the area to be measured;

[0216] Constructing a sample set from the historical radar echo graphs before a certain moment and the historical radar echo graphs after the certain moment;

[0217] Dividing the sample set into a training set and a test set according to a set ratio;

[0218] The historical radar echo map before the certain moment is used as input, and the historical radar echo map after the certain moment is used as output to train the dual-channel convolution time series extrapolation model, so as to obtain a preliminarily trained dual-channel convolution time series extrapolation model;

[0219] The preliminarily trained dual-channel convolution timing extrapolation model is tested based on the test set, and the preliminarily trained dual-channel convolution timing extrapolation model that passes the test is used as the trained dual-channel convolution timing extrapolation model.

[0220] Optionally, building blocks are also included for:

[0221] An encoder is constructed by a time sub-network for extracting time data, a space sub-network for extracting space data, and a spatiotemporal data fusion module for fusing the time data with the space data, wherein the time sub-network and the space sub-network are in parallel;

[0222] A dual-channel convolutional temporal extrapolation model is constructed by an encoder and a decoder that decodes high-bit fused features into predicted future spatiotemporal data.

[0223] Example 4

[0224] like Figure 8 As shown, the present invention also provides an electronic device, which may be a computer device, a single-chip device, an intelligent mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, the processor, and the transceiver component are connected via a bus; the memory may be used to store an execution program, and an exemplary execution program may include instructions; the processor is used to execute the instructions stored in the memory. The memory may also be used to store data, which may be called and / or modified when the instructions are executed.

[0225] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is suitable for implementing one or more instructions, and is specifically suitable for loading and executing one or more instructions in a storage medium to implement a corresponding method flow or a corresponding function, so as to implement the steps of a nonlinear short-term prediction method based on radar data in the above-mentioned embodiment.

[0226] Example 5

[0227] Based on the same inventive concept, the present invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory), which is a memory device in the electronic device for storing programs and data. It can be understood that the storage medium here can include both the built-in storage medium in the electronic device and the extended storage medium supported by the electronic device. The storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor are also stored in the storage space, and these instructions can be one or more execution programs (including program codes). It should be noted that the storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. The processor loads and executes one or more instructions stored in the storage medium, which can implement the steps of a nonlinear short-term prediction method based on radar data in the above embodiment.

[0228] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0229] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0230] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0231] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0232] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are included in the scope of the claims of the present invention to be approved.

Claims

1. A nonlinear short-term forecasting method based on radar data, characterized in that: include: Acquire historical radar echo data before the time to be predicted, and preprocess the radar echo data to obtain a radar echo map; Substituting the radar echo map into a pre-trained dual-channel convolutional time series extrapolation model to obtain radar extrapolation prediction data; The pre-trained dual-channel convolutional time series extrapolation model is obtained by learning the radar echo map before a certain moment in history and the radar echo map after the certain moment by using deep learning convolution; The dual-channel convolutional timing extrapolation model includes a deep learning convolutional dual-channel encoder and decoder.

2. The method according to claim 1, characterized in that The acquiring of historical radar echo data before the time to be predicted and preprocessing of the radar echo data to obtain a radar echo map includes: Acquire radar echo data within a preset period before the time to be predicted; Extracting information from the radar echo data to obtain the latitude and longitude range of the area covered by the radar echo; The longitude and latitude range of the coverage area of ​​the radar echo is clipped according to the set spatial resolution to obtain a clipped radar echo map; The data in the clipped radar echo map is quality screened to obtain a high-quality radar echo map.

3. The method according to claim 2, characterized in that The extracting information from the radar echo data to obtain the latitude and longitude range of the coverage area of ​​the radar echo includes: Calculate the longitude and latitude coordinates of each grid point from the data area description parameters in the file header of the radar echo data; Based on the description of the radar file in the file header, obtain the latitude and longitude range of the area covered by the radar echo.

4. The method according to claim 2, characterized in that The longitude and latitude range of the coverage area of ​​the radar echo is clipped according to the set spatial resolution to obtain a clipped radar echo map, including: Determine the core area of ​​current nowcast modeling based on the latitude and longitude range of the coverage area of ​​the radar echo; The radar echo map is cropped with the central longitude and latitude of the core area as the center and with the set spatial resolution to obtain a cropped radar echo map.

5. The method according to claim 2, characterized in that The quality screening of the data in the clipped radar echo map to obtain a high-quality radar echo map includes: The points whose radar echo values ​​in the clipped radar echo map satisfy the first relational expression are regarded as isolated high values; The points in the clipped radar echo map whose radar echo values ​​satisfy the second relational expression are regarded as ground object echo points; Isolated high values ​​and ground object echo points are removed from the clipped radar echo map to obtain a high-quality radar echo map.

6. The method according to claim 5, characterized in that The first relationship is shown as follows: Where M 5*5 is a 5×5 grid range within the surrounding area, M i The 5×5 grid range M within the surrounding area 5*5 The radar echo value of any point except the center point within x,y is a point in the radar echo graph.

7. The method according to claim 5, characterized in that The second relationship is as follows: Where N 3*3 is a 3×3 grid range within the surrounding area, N i The 3×3 grid range N within the surrounding area 3*3 The radar echo value of any point within.

8. The method according to claim 1, characterized in that The training of the dual-channel convolutional timing extrapolation model includes: Obtain the historical radar echo map of the area to be measured; Constructing a sample set from the historical radar echo graphs before a certain moment and the historical radar echo graphs after the certain moment; Dividing the sample set into a training set and a test set according to a set ratio; The historical radar echo map before the certain moment is used as input, and the historical radar echo map after the certain moment is used as output to train the dual-channel convolution time series extrapolation model, so as to obtain a preliminarily trained dual-channel convolution time series extrapolation model; The preliminarily trained dual-channel convolution timing extrapolation model is tested based on the test set, and the preliminarily trained dual-channel convolution timing extrapolation model that passes the test is used as the trained dual-channel convolution timing extrapolation model.

9. The method according to claim 8, characterized in that The step of testing the preliminarily trained dual-channel convolution timing extrapolation model based on the test set, and using the preliminarily trained dual-channel convolution timing extrapolation model that passes the test as a trained dual-channel convolution timing extrapolation model, includes: Substituting the historical radar echo map before a certain moment in the test set into the preliminarily trained dual-channel convolutional time series extrapolation model to obtain the radar extrapolation prediction data at the said certain moment; Calculate the evaluation index value based on the radar extrapolation prediction data at the certain moment combined with the historical radar data after a certain moment in the test set; Based on whether the evaluation index value meets a preset threshold, if so, the test is passed, and the preliminarily trained dual-channel convolution timing extrapolation model is used as the trained dual-channel convolution timing extrapolation model; otherwise, the test is not passed; The evaluation index values ​​include: correlation coefficient, root mean square error, and image structure similarity index.

10. The method according to claim 1, characterized in that The construction of the dual-channel convolutional timing extrapolation model includes: An encoder is constructed by a time sub-network for extracting time data, a space sub-network for extracting space data, and a spatiotemporal data fusion module for fusing the time data with the space data, wherein the time sub-network and the space sub-network are in parallel; A dual-channel convolutional temporal extrapolation model is constructed by an encoder and a decoder that decodes high-bit fused features into predicted future spatiotemporal data.

11. The method according to claim 10, characterized in that Substituting the radar echo map into a pre-trained dual-channel convolutional time series extrapolation model to obtain radar extrapolation prediction data includes: extracting time features and space features from the processed radar echo data through a time sub-network and a space sub-network; The temporal feature and the spatial feature are fused by a spatiotemporal data fusion module to obtain a fused feature; The fused features are decoded by a decoder, and the fused features are converted into radar extrapolation prediction data.

12. A nonlinear short-term forecasting device based on radar data, characterized in that: include: A parameter processing module, used for acquiring historical radar echo data before the time to be predicted, and preprocessing the radar echo data to obtain a radar echo map; A prediction module, used for substituting the radar echogram into a pre-trained dual-channel convolution time series extrapolation model to obtain radar extrapolation prediction data; The pre-trained dual-channel convolutional time series extrapolation model is obtained by learning the radar echo map before a certain moment in history and the radar echo map after the certain moment by using deep learning convolution; The dual-channel convolutional timing extrapolation model includes a deep learning convolutional dual-channel encoder and decoder.

13. The device according to claim 12, characterized in that The parameter processing module comprises: The acquisition submodule is used to acquire radar echo data within a preset period before the time to be predicted; An extraction submodule, used to extract information from the radar echo data to obtain the latitude and longitude range of the coverage area of ​​the radar echo; A clipping submodule, used to clip the latitude and longitude range of the coverage area of ​​the radar echo according to a set spatial resolution to obtain a clipped radar echo map; The screening submodule is used to perform quality screening on the data in the clipped radar echo map to obtain a high-quality radar echo map.

14. The device according to claim 13, characterized in that The extraction submodule is specifically used for: Calculate the longitude and latitude coordinates of each grid point from the data area description parameters in the file header of the radar echo data; Based on the description of the radar file in the file header, obtain the latitude and longitude range of the area covered by the radar echo.

15. The device according to claim 13, characterized in that The cutting submodule is specifically used for: Determine the core area of ​​current nowcast modeling based on the latitude and longitude range of the coverage area of ​​the radar echo; The radar echo map is cropped with the central longitude and latitude of the core area as the center and with the set spatial resolution to obtain a cropped radar echo map.

16. The device according to claim 13, characterized in that The screening submodule is specifically used for: The points whose radar echo values ​​in the clipped radar echo map satisfy the first relational expression are regarded as isolated high values; The points in the clipped radar echo map whose radar echo values ​​satisfy the second relational expression are regarded as ground object echo points; Isolated high values ​​and ground object echo points are removed from the clipped radar echo map to obtain a high-quality radar echo map.

17. The device according to claim 16, characterized in that The first relationship is shown as follows: Where M 5*5 is a 5×5 grid range within the surrounding area, M i The 5×5 grid range M within the surrounding area 5*5 The radar echo value of any point except the center point within x,y is a point in the radar echo graph.

18. The device according to claim 16, characterized in that The second relationship is as follows: Where N 3*3 is a 3×3 grid range within the surrounding area, N i The 3×3 grid range N within the surrounding area 3*3 The radar echo value of any point within.

19. The device according to claim 12, characterized in that Also includes training modules for: Obtain the historical radar echo map of the area to be measured; Constructing a sample set from the historical radar echo graphs before a certain moment and the historical radar echo graphs after the certain moment; Dividing the sample set into a training set and a test set according to a set ratio; The historical radar echo map before the certain moment is used as input, and the historical radar echo map after the certain moment is used as output to train the dual-channel convolution time series extrapolation model, so as to obtain a preliminarily trained dual-channel convolution time series extrapolation model; The preliminarily trained dual-channel convolution timing extrapolation model is tested based on the test set, and the preliminarily trained dual-channel convolution timing extrapolation model that passes the test is used as the trained dual-channel convolution timing extrapolation model.

20. The device according to claim 12, characterized in that Also included are building blocks for: An encoder is constructed by a time sub-network for extracting time data, a space sub-network for extracting space data, and a spatiotemporal data fusion module for fusing the time data with the space data, wherein the time sub-network and the space sub-network are in parallel; A dual-channel convolutional temporal extrapolation model is constructed by an encoder and a decoder that decodes high-bit fused features into predicted future spatiotemporal data.

21. An electronic device, characterized in that: include: at least one processor and memory; The memory and the processor are connected via a bus; The memory is used to store one or more programs; When the one or more programs are executed by the at least one processor, a nonlinear short-term forecasting method based on radar data as claimed in any one of claims 1 to 11 is implemented.

22. A readable storage medium, characterized in that: An execution program is stored thereon, and when the execution program is executed, a nonlinear short-term prediction method based on radar data as described in any one of claims 1 to 11 is implemented.

Images