High-temporal-spatial-resolution weather radar beam propagation path inversion method

Abstract

The invention relates to a high-temporal-spatial-resolution weather radar beam propagation path inversion method, which belongs to the field of radar detection, and designs a more accurate temporal-spatial matching strategy by taking meteorological element data output by WRF as a background field and combining high-frequency radio sounding data. A neural network model capable of fully mining local and global spatial features of data is constructed, a long-time and short-time feature extraction and fusion model is constructed, efficient and accurate extraction of time and spatial features of different scales of meteorological elements is realized, an input data set time and spatial complexity calculation method is further designed, and the calculation efficiency is improved. The adaptive optimization selection of the convolution kernel is realized, the real-time performance of outputting high-temporal-spatial-resolution meteorological data based on the 3D temporal-spatial convolution model is greatly improved, the atmospheric refractive index model and the electromagnetic wave propagation path calculation model are introduced, and the inversion of the high-temporal-spatial-resolution and accurate weather radar beam propagation path is realized.

Description

Technical Field

[0001] The present invention relates to the field of radar detection, and in particular to a high temporal and spatial resolution weather radar beam propagation path inversion method. Background Art

[0002] The accurate calculation of the weather radar beam propagation path is the basis for determining the effective monitoring range of the radar. Since the propagation of the weather radar beam is affected by factors such as terrain, meteorological conditions (such as temperature, humidity and wind speed), beam characteristics (such as beam width and wavelength), and atmospheric refraction, understanding the beam propagation path can help determine the effective distance and area of ​​radar detection. For example, radar signals may be affected by atmospheric refraction, terrain shielding and other factors, resulting in some areas being unable to be effectively monitored. In addition, the calculation of the weather radar beam propagation path can help evaluate the performance of the radar system and provide data support for the design and optimization of the radar system. For example, the analysis of the beam propagation path helps to determine the best radar installation position, the best antenna height, the beam tilt angle and other parameters, thereby optimizing the detection range and accuracy of the radar system. In addition, in a multi-radar system, the calculation of the beam propagation path helps to understand the overlap and blind spots of the observation areas between multiple radars. By optimizing the installation position and beam propagation path of the radar, the coverage efficiency and collaborative observation capability of the radar networking system can be improved, which is especially important in large-scale weather monitoring and early warning.

[0003] At present, there are several common methods for obtaining weather radar beam propagation paths: (1) Standard atmospheric model method. This method assumes that the atmospheric conditions conform to the standard atmospheric model. It is a relatively simple calculation method and is suitable for beam propagation path calculation under general circumstances. However, this method ignores the changes in actual atmospheric conditions and cannot accurately describe the impact of complex weather phenomena such as temperature inversion and air pressure fluctuations on beam propagation; (2) Atmospheric refraction model method. When calculating the beam propagation path, this method focuses on the impact of atmospheric refraction on the beam and can more accurately simulate the beam propagation path under different meteorological conditions. This method requires high-precision atmospheric data and complex calculation processes. In particular, real-time calculation is difficult under dynamically changing meteorological conditions; (3) Numerical weather forecast model method. This method combines various meteorological parameters in the atmosphere (such as temperature, humidity, air pressure, wind speed, etc.) and solves the radar beam propagation path through numerical simulation. It can consider the interaction of multiple dynamic factors in the atmosphere, such as wind field changes, temperature and humidity changes, and is suitable for complex meteorological environments. However, the disadvantage is that it requires a large amount of computing resources and high-precision meteorological data, and its real-time performance is poor. (4) Terrain influence model method: When calculating the propagation path of the radar beam, the terrain influence model needs to consider the refraction and blocking effect of the terrain on the beam. This method uses digital elevation model (DEM) data to model terrain obstacles and refraction effects. The advantage is that it can simulate the propagation of the beam under complex terrain with high accuracy. The disadvantage is that it requires detailed terrain data and has a large amount of calculation.

[0004] Although the above methods have their own advantages and disadvantages, in general, it is impossible to achieve accurate inversion of weather radar beam propagation paths with high spatial resolution and real-time by relying solely on the above single methods. The root cause is that it is difficult to accurately obtain meteorological element data with high temporal and spatial resolution. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a high temporal and spatial resolution weather radar beam propagation path inversion method to solve the deficiencies of the prior art.

[0006] The object of the present invention is achieved by the following technical solution: a high temporal and spatial resolution weather radar beam propagation path inversion method, the method comprising:

[0007] S1. Input the collected data into geographic information processing software, calculate the data set A according to the collected data, calculate the spatial distance difference, spatial distance weight value, temporal distance difference and temporal distance weight value according to the data set A, and construct the spatiotemporal distance weight, and use the minimum value of the spatiotemporal distance weight as the screening condition to screen the data set A, and obtain the WRF atmospheric temperature data, WRF atmospheric humidity data and WRF atmospheric pressure data that match the spatiotemporal data;

[0008] S2, the data generated by S1 are used as input data set 1, the high-frequency radio sounding meteorological data collected by S1 are used as output data set, a multi-scale convolutional neural network is constructed based on the original model of the convolutional neural network, and finally the local spatial feature weight coefficients of atmospheric temperature, atmospheric humidity and atmospheric pressure are obtained, a dynamic weight generation model is constructed based on a small fully connected neural network model, and finally the spatial feature vector after weighted fusion of atmospheric temperature, atmospheric humidity and atmospheric pressure is obtained, and a neural network constructed by a C-layer 1D convolutional network and a bidirectional LSTM model are constructed, and finally the time feature vector of atmospheric temperature, atmospheric humidity and atmospheric pressure data is obtained;

[0009] S3, the spatial feature vector and time feature vector after weighted fusion of meteorological data generated by S2, and the data generated by S1 form input data set 2, and calculate its comprehensive complexity, take the high-frequency radiosonde meteorological data collected by S1 as the output data set, introduce the 3D spatiotemporal convolutional network model to construct a 3D spatiotemporal convolutional model of atmospheric temperature, atmospheric humidity and atmospheric pressure;

[0010] S4, input the spatial feature vector and time feature vector after weighted fusion of the meteorological data generated by S2, and the data generated by S1 into the joint model, and the generated data are respectively input into the 3D spatiotemporal convolution model of atmospheric temperature, atmospheric humidity and atmospheric pressure generated by S3 to obtain atmospheric temperature data, atmospheric humidity data and atmospheric pressure data with high spatiotemporal resolution;

[0011] S5. Connect the two calculation models of atmospheric refractive index and electromagnetic wave propagation path in series, input the data obtained by S4 and the weather radar geographic location information and antenna operation parameter information data collected by S1 into the series model, and realize the accurate inversion of weather radar beam propagation path with high temporal and spatial resolution.

[0012] The data collected in S1 include:

[0013] Collect meteorological data and auxiliary data output by mesoscale WRF at multiple times. The meteorological data output by WRF includes atmospheric temperature, atmospheric humidity and atmospheric pressure. The auxiliary data output by WRF includes the latitude, longitude and altitude of the meteorological data, temporal resolution, spatial resolution and the time corresponding to the data. WRF is a weather forecast mode.

[0014] Collect high-frequency radiosonde meteorological data and auxiliary data at multiple times. High-frequency radiosonde meteorological data include atmospheric temperature, atmospheric humidity and atmospheric pressure. High-frequency radiosonde auxiliary data include the latitude, longitude and altitude of the data, time resolution, spatial resolution and the time corresponding to the data.

[0015] Collect weather radar geographic location information and antenna operating parameter information, including the longitude, latitude, altitude, antenna elevation angle and antenna azimuth angle of the weather radar site.

[0016] The S1 specifically includes the following contents:

[0017] A1. Input the collected atmospheric temperature data and auxiliary data output by WRF at multiple times into the geographic information processing software, set the spatial interval of the spatial resolution grid to ΔS, select the bilinear interpolation scheme, and run the geographic information processing software to generate WRF atmospheric temperature data with medium spatial resolution, recorded as T_WRF_M, and its time information is recorded as , the spatial position consists of longitude, latitude and altitude, recorded as ;

[0018] A2. The atmospheric temperature data from the high-frequency radiosonde meteorological data collected at multiple times is recorded as T_Radio, and its time information is recorded as , the spatial position consists of longitude, latitude and altitude, recorded as ;

[0019] A3. Set the height difference threshold of the adjacent spaces at the same altitude level to , calculate the absolute height difference between a data point P in T_Radio and all T_WRF_M data points , select the height difference less than The corresponding T_WRF_M data points are recorded as data set A;

[0020] A4. Calculate the spatial distance difference between point P and the data points in data set A according to the spatial distance difference calculation formula , calculate the spatial distance weight value from the data point in data set A to point P according to the spatial distance weight calculation formula ;

[0021] A5. Calculate the time distance difference between point P and the data point in data set A based on the time data of T_WRF_M and T_Radio. , calculate the time distance weight value from the data point in data set A to point P according to the time distance weight calculation formula ;

[0022] A6. Constructing space-time distance weights ,by The minimum value is used as the screening condition, and the data points in the screened data set A are used as the spatiotemporal matching points of point P;

[0023] A7. Repeat steps A3-A6 to process all data points of T_Radio one by one to obtain the WRF atmospheric temperature dataset that matches the T_Radio data in time and space, recorded as Tm_WRF_M.

[0024] A8. Apply methods A2-A7 to the collected WRF atmospheric humidity and atmospheric pressure data to obtain WRF atmospheric humidity data Qm_WRF_M and atmospheric pressure data Pm_WRF_M that are temporally and spatially matched with the high-frequency radiosonde meteorological data.

[0025] The S2 specifically includes the following contents:

[0026] B1. The generated Tm_WRF_M and the corresponding altitude, longitude, latitude, and time data are used as the first input data set. The atmospheric temperature data T_Radio in the collected high-frequency radiosonde meteorological data is used as the output data set. A multi-scale convolutional neural network is constructed based on the original convolutional neural network model. The local spatial feature weight coefficients of atmospheric temperature, atmospheric humidity, and atmospheric pressure are obtained through the multi-scale convolutional neural network.

[0027] B2. The generated Tm_WRF_M and the corresponding altitude, longitude, latitude and time data are used as the second input data set. The atmospheric temperature data T_Radio in the collected high-frequency radiosonde meteorological data is used as the output data set. A dynamic weight generation model is constructed based on a small fully connected neural network model. The weighted spatial characteristic vector of atmospheric temperature, atmospheric humidity and atmospheric pressure is obtained according to the dynamic weight generation model.

[0028] B3. The collected data is processed by S1 to obtain Tm_WRF_M and T_Radio at multiple times and the corresponding altitude, longitude, latitude, and time data to form the third input data set. T_Radio at multiple times is used as the output data set. A neural network constructed by a C-layer 1D convolutional network and a bidirectional LSTM model are constructed to generate the time feature vectors of atmospheric temperature, atmospheric humidity, and atmospheric pressure data.

[0029] The B1 specifically includes the following contents:

[0030] B11, the generated Tm_WRF_M and the corresponding altitude, longitude, latitude and time data are used as the first input data set, and the atmospheric temperature data T_Radio in the high-frequency radiosonde meteorological data collected at multiple times are used as the output data set;

[0031] B12. Based on the original model of convolutional neural network, K convolution kernels of different sizes are set, and the dynamic tuning strategy based on evolution is selected to construct a multi-scale convolutional neural network. Each scale of convolution kernel constitutes a convolution channel separately, and convolution operations are performed on the input data set respectively to extract the spatial features of different scales of the input data set, which is recorded as ,in Respectively represent the longitude, latitude, and altitude of the data point, k=1,2,3,…,K;

[0032] B13. Calculate based on the atmospheric temperature gradient change rate formula and the mean square error formula The gradient change rate and mean square error , the normalized gradient change rate and mean square error are processed using the normalization formula to obtain the normalized gradient change rate and mean square error ;

[0033] B14. Calculate the local spatial characteristic weight coefficient of atmospheric temperature using convolution kernels of different sizes according to the local weight formula ;

[0034] B15. Apply the methods of B11-B14 to the generated Qm_WRF_M data and Pm_WRF_M data to obtain the local spatial characteristic weight coefficients of atmospheric humidity and atmospheric pressure respectively. and .

[0035] The B2 specifically includes the following contents:

[0036] B21, the generated Tm_WRF_M and the corresponding altitude, longitude, latitude, and time data are used as the second input data set, and the atmospheric temperature data T_Radio in the high-frequency radiosonde meteorological data collected at multiple times are used as the output data set;

[0037] B22. Based on a small fully connected neural network model, K convolution kernels of different sizes are set. The global average feature extraction formula is used to process each type of data in the input data set to obtain the global feature ;

[0038] B23. Random sampling is performed in the Gaussian distribution sequence to generate the initial matrix of the connection layer weights, the loss function is set to the mean square error between the input and the input data set, the gradient descent strategy is selected to iteratively optimize the model performance, and a dynamic weight generation model for global features is constructed. The model is run to obtain the global feature weight coefficients of convolution kernels of different scales. ;

[0039] B24. According to the local spatial feature weight coefficients of convolution kernels of different sizes, a local and global feature weight coefficient balance model is constructed, and the global feature weight coefficient is corrected to obtain the corrected weight coefficient. ;

[0040] B25, according to Spatial features extracted by convolution of K different scales Perform weighted fusion processing to obtain the spatial characteristics of atmospheric temperature after weighted fusion ;

[0041] B26. Apply the methods of B21-B25 to the generated Qm_WRF_M data and Pm_WRF_M data to obtain the weighted fusion spatial feature vectors of atmospheric humidity and atmospheric pressure, respectively. and .

[0042] The B3 specifically includes the following contents:

[0043] B31, after processing the collected WRF meteorological data and auxiliary data at multiple times, as well as the high-frequency radiosonde meteorological data and auxiliary data, obtain Tm_WRF_M, T_Radio and corresponding altitude, longitude, latitude, and time data at multiple times, use Tm_WRF_M at multiple times and the corresponding altitude, longitude, latitude, and time data as the third input data set, and use T_Radio at multiple times as the output data set;

[0044] B32. Construct a neural network consisting of C layers of 1D convolutional networks. Each layer of the convolutional network uses a different small-scale convolution kernel to perform convolution operations on the input data set and extract C types of short-time scale change features, recorded as , ,…, ;

[0045] B33. Set L sliding windows with longer time scales, and record the window sizes as Long_1, Long_2, …, Long_L respectively. Build a bidirectional LSTM model, set the number of LSTM units as NUM, and run the bidirectional LSTM model under different sliding window length conditions to obtain L long time scale change characteristics, recorded as , ,…, ;

[0046] B34. Combine the short-time scale variation characteristics with the long-time scale variation characteristics to obtain the time characteristic vector of the atmospheric temperature data in the third input data set. ;

[0047] B35. Apply the methods of B31-B34 to the Qm_WRF_M data and Pm_WRF_M data at multiple times to obtain the time characteristic vectors of atmospheric humidity and atmospheric pressure data respectively. and .

[0048] The S3 specifically includes the following contents:

[0049] C1, generate , , and the extracted Tm_WRF_M together constitute the input data set 2. The spatial complexity of the input data set is calculated according to the spatial autocorrelation calculation formula. , the time complexity of the input data set is calculated according to the time autocorrelation calculation formula , add the time complexity and space complexity to get the comprehensive complexity ;

[0050] C2, the atmospheric temperature data from the high-frequency radiosonde data collected at multiple times are used as the output data set, and the 3D spatiotemporal convolutional network model is introduced. To select the convolution kernel, train the network model, and build a 3D spatiotemporal convolution model of atmospheric temperature;

[0051] C3. Apply the methods of C1 and C2 to the atmospheric humidity and atmospheric pressure data to obtain the 3D convolution models of atmospheric humidity and atmospheric pressure, respectively.

[0052] The S4 specifically includes the following contents:

[0053] D1, generate , , and the extracted Tm_WRF_M are input into a joint model consisting of a deconvolution model and a temporal super-resolution model, and the joint model is run to generate a high spatial resolution feature vector of atmospheric temperature. , high temporal resolution feature vector , and Tm_WRF_H with high temporal and spatial resolution;

[0054] D2. Apply the method of D1 to the atmospheric humidity and atmospheric pressure data to obtain the high spatial resolution feature vectors of atmospheric humidity and atmospheric pressure respectively. , , a high temporal resolution feature vector , , and Qm_WRF_H and Pm_WRF_H with high temporal and spatial resolution;

[0055] D3. , The data are input together with Tm_WRF_H into the atmospheric temperature 3D spatiotemporal convolution model, and the model is run to generate atmospheric temperature data with high spatiotemporal resolution;

[0056] D4. , The data are input into the atmospheric humidity 3D spatiotemporal convolution model together with Qm_WRF_H, and the model is run to generate atmospheric humidity data with high spatiotemporal resolution;

[0057] D5. , The data are input together with Pm_WRF_H into the 3D spatiotemporal convolution model of atmospheric pressure, and the model is run to generate atmospheric pressure data with high spatiotemporal resolution.

[0058] The present invention has the following advantages: a high spatiotemporal resolution weather radar beam propagation path inversion method, which realizes the accurate spatiotemporal matching of medium spatial resolution WRF meteorological data and radiosonde data; realizes the extraction of spatial features of different scales, and then realizes the calculation of local spatial feature weight coefficients; realizes the calculation of global feature weight coefficients; realizes the correction of global feature weight coefficients, and generates weighted fusion spatial feature vectors of atmospheric temperature, humidity and air pressure; realizes the extraction of multiple short-time scale and long-time scale features; realizes the generation of high spatiotemporal resolution atmospheric temperature, humidity and air pressure data; realizes the inversion of high spatiotemporal resolution and accurate weather radar beam propagation path. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a schematic diagram of the process of the present invention;

[0060] Figure 2 It is a flow chart for realizing the spatiotemporal matching of WRF meteorological data and radiosonde data of the present invention;

[0061] Figure 3 It is a flowchart for realizing the calculation of the weight coefficient of the local space feature of the present invention;

[0062] Figure 4 It is a flowchart for realizing the calculation of the global feature weight coefficient of the present invention;

[0063] Figure 5 A flow chart for realizing the generation of weighted fusion spatial feature vectors of meteorological data of the present invention;

[0064] Figure 6 It is a flow chart for realizing the fusion of long- and short-time scale features of meteorological data of the present invention;

[0065] Figure 7 A flowchart for realizing the generation of high temporal and spatial resolution meteorological data of the present invention;

[0066] Figure 8 The present invention is a flowchart for implementing the high spatiotemporal resolution weather radar beam propagation path inversion. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application provided below in conjunction with the drawings is not intended to limit the scope of protection of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application. The present invention is further described below in conjunction with the drawings.

[0068] The present invention specifically relates to a high-temporal-spatial-resolution weather radar beam propagation path inversion method. The meteorological element data output by WRF is used as the background field, combined with high-frequency radio sounding data, a more accurate time-space matching strategy is designed, a neural network model that can fully mine the local and global spatial characteristics of the data is constructed, a long- and short-time feature extraction and fusion model is constructed, and the efficient and accurate extraction of time and space characteristics of meteorological elements at different scales is achieved. Then, a method for calculating the time and space complexity of the input data set is designed, and the adaptive optimization selection of the convolution kernel is achieved, which greatly improves the real-time performance of the high-temporal-spatial-resolution meteorological data output based on the 3D space-time convolution model, introduces an atmospheric refractive index model and an electromagnetic wave propagation path calculation model, and achieves high-temporal-spatial-resolution and accurate inversion of the weather radar beam propagation path.

[0069] like Figure 1 As shown, specifically including the following:

[0070] Step 1: Collect meteorological data and auxiliary data output by the mesoscale weather forecast model (WRF) at multiple times; collect high-frequency radiosonde meteorological data and auxiliary data at multiple times; collect weather radar geographic location information and antenna operating parameter information;

[0071] Among them, the meteorological data output by WRF mainly include atmospheric temperature, atmospheric humidity, air pressure and other data;

[0072] The auxiliary data output by WRF mainly includes the longitude and latitude and altitude of the meteorological data, the temporal resolution, spatial resolution and the time corresponding to the data;

[0073] High-frequency radiosonde meteorological data sets mainly include atmospheric temperature, atmospheric humidity, atmospheric pressure and other data;

[0074] The high-frequency radiosonde auxiliary data set mainly includes data such as the latitude and longitude of the data, the observation time interval and height, and the observation time;

[0075] The weather radar geographical location information and antenna operating parameter information mainly include the longitude, latitude, altitude, antenna elevation angle, antenna azimuth angle, etc. of the weather radar site;

[0076] Step 2: If Figure 2 As shown in Figure 1, first, input the atmospheric temperature data and auxiliary data of WRF output at multiple moments collected in step 1 into the geographic information processing software, set the spatial interval of the spatial resolution grid to 3 kilometers, select the bilinear interpolation scheme, and run the software to generate WRF atmospheric temperature data with medium spatial resolution, denoted as T_WRF_M, and its time information is recorded as , the spatial position consists of longitude, latitude and altitude, recorded as ; Secondly, the atmospheric temperature data in the high-frequency radiosonde meteorological data collected at multiple times in step 1 is recorded as T_Radio, and its time information is recorded as , the spatial position consists of longitude, latitude and altitude, recorded as ; Then, the height difference threshold in the adjacent space at the same altitude layer is set to = 50 meters, calculate the absolute height difference between a data point in T_Radio (denoted as point P) and all T_WRF_M data points , select the height difference less than The corresponding T_WRF_M data point is recorded as data set A; then, the spatial distance difference between point P and the data point in data set A is calculated according to the spatial distance difference calculation formula. , calculate the spatial distance weight value from the data point in data set A to point P according to the spatial distance weight calculation formula ; Then, based on the time data of T_WRF_M and T_Radio, calculate the time distance difference between point P and the data points in data set A , calculate the time distance weight value from the data point in data set A to point P according to the time distance weight calculation formula Finally, construct the space-time distance weight ,by The minimum value is used as the screening condition, and the data points in the screened data set A are used as the spatiotemporal matching points of point P. According to this method, all data points of T_Radio are processed one by one to obtain the WRF atmospheric temperature data set that is spatiotemporally matched with the T_Radio data, which is recorded as Tm_WRF_M. The above scheme is applied to the WRF atmospheric humidity and atmospheric pressure data collected in step 1, respectively, to obtain the WRF atmospheric humidity data Qm_WRF_M and atmospheric pressure data Pm_WRF_M that are spatiotemporally matched with the high-frequency radiosonde meteorological data.

[0077] The calculation formula of spatial distance difference is: ;

[0078] The time distance difference calculation formula is: ;

[0079] The spatial distance weight calculation formula is: ;

[0080] The time distance weight calculation formula is: ;

[0081] Step 3: If Figure 3 As shown in the figure, the Tm_WRF_M generated in step 2 and the corresponding altitude, longitude, latitude, time data, etc. are used as the input data set, and the atmospheric temperature data T_Radio in the high-frequency radiosonde meteorological data collected in step 1 is used as the output data set. Based on the original model of the convolutional neural network, three convolution kernels of different sizes are set, namely 3×3, 5×5, and 7×7. The dynamic tuning strategy based on evolution is selected to construct a multi-scale convolutional neural network. The convolution kernel of each scale constitutes a convolution channel separately, and the convolution operation is performed on the input data set respectively to realize the extraction of spatial features of different scales of the input data set, which is recorded as ,in Represents the longitude, latitude, and altitude of the data point, k=1,2,3; then, calculate according to the atmospheric temperature gradient change rate formula and the mean square error formula The gradient change rate and mean square error , the normalized gradient change rate and mean square error are processed using the normalization formula to obtain the normalized gradient change rate and mean square error Finally, the local spatial characteristic weight coefficient of atmospheric temperature obtained by convolution kernels of different sizes is calculated according to the local weight formula. Applying the above scheme to the Qm_WRF_M data and Pm_WRF_M data generated in step 2, the local spatial characteristic weight coefficients of atmospheric humidity and atmospheric pressure can be obtained respectively. and .

[0082] Among them, the gradient change rate formula is: ;

[0083] The mean square error formula is: ,in for The mean of

[0084] The normalization formula is: ,in is the variable to be normalized;

[0085] The local weight formula is: ,in and are the weight hyperparameters for the gradient change rate and mean square error respectively;

[0086] Step 4: Figure 4 and Figure 5 As shown in the figure, the Tm_WRF_M generated in step 2 and the corresponding altitude, longitude, latitude, time data, etc. are used as the input data set, and the atmospheric temperature data T_Radio in the multiple high-frequency radiosonde meteorological data collected in step 1 is used as the output data set. Based on the small fully connected neural network model, the same three convolution kernels of different sizes as in step 3 are set. For each type of data in the input data set, the global average feature extraction formula is used for processing to obtain the global feature After that, random sampling is performed from the Gaussian distribution sequence to generate the initial matrix of the connection layer weights, the loss function is set to the mean square error between the input and the input data set, the gradient descent strategy is selected to iteratively optimize the model performance, and a dynamic weight generation model for global features is constructed. Running the model can obtain the global feature weight coefficients of convolution kernels of different scales. After that, the local spatial feature weight coefficients of the convolution kernels of different sizes obtained in step 3 are combined to construct a balance model of local and global feature weight coefficients, and then the global feature weight coefficients obtained in step 4 are corrected to obtain the corrected weight coefficients Finally, according to Spatial features extracted by convolution of K different scales Perform weighted fusion processing to obtain the spatial characteristics of atmospheric temperature after weighted fusion Applying the above scheme to the Qm_WRF_M data and Pm_WRF_M data generated in step 2, the weighted fusion spatial feature vectors of atmospheric humidity and atmospheric pressure can be obtained respectively. and .

[0087] Among them, the global average feature extraction formula is: ,in For input data;

[0088] The local and global feature weight coefficient balance model is: ;

[0089] The weighted fusion process is: ;

[0090] Step 5: Figure 5 and Figure 6 As shown in the figure, after the WRF meteorological data and auxiliary data at multiple times collected in step 1, as well as the high-frequency radiosonde meteorological data and auxiliary data are processed in step 2, Tm_WRF_M, T_Radio at multiple times and the corresponding data such as altitude, longitude, latitude, time, etc. are obtained. Tm_WRF_M at multiple times and the corresponding data such as altitude, longitude, latitude, time, etc. are used as input data sets, and T_Radio at multiple times is used as output data sets. A neural network consisting of 2 layers of 1D convolutional networks is constructed. Each layer of the convolutional network uses different small-scale convolution kernels to perform convolution operations on the input data sets respectively, and extracts the change characteristics of two short time scales of 3 hours and 5 hours, which are recorded as , Similarly, Tm_WRF_M at multiple times and the corresponding altitude, longitude, latitude, time and other data are used as input data sets, and T_Radio at multiple times is used as output data sets. Sliding windows of 6 hours, 12 hours and 24 hours and three longer time scales are set. The window sizes are recorded as Long_1, Long_2, and Long_3 respectively. A bidirectional LSTM model is constructed, and the number of LSTM units is set to 128. Under different sliding window length conditions, running the model separately can obtain three long time scale change characteristics, which are recorded as , , After that, the short-time scale variation characteristics are fused with the long-time scale variation characteristics to obtain the time characteristic vector of the atmospheric temperature data in the input data set. Applying the above scheme to the Qm_WRF_M data and Pm_WRF_M data at multiple times, the time characteristic vectors of atmospheric humidity and atmospheric pressure data can be obtained respectively. and .

[0091] Step 6: Figure 7 As shown, the generated in step 4 , generated in step 5 , and Tm_WRF_M extracted in step 2 together constitute the input data set, and the spatial complexity of the input data set is calculated according to the spatial autocorrelation calculation formula . Calculate the time complexity of the input data set according to the time autocorrelation calculation formula ; Add the time complexity and space complexity to get the comprehensive complexity Afterwards, the atmospheric temperature data in the high-frequency radiosonde data collected in step 1 is used as the output data set and introduced into the 3D spatiotemporal convolutional network model. By selecting the convolution kernel of appropriate size and training the network model, a 3D spatiotemporal convolution model of atmospheric temperature can be constructed. By applying the above scheme to atmospheric humidity and atmospheric pressure data, 3D convolution models of atmospheric humidity and atmospheric pressure can be obtained respectively.

[0092] The spatial autocorrelation calculation formula is: ,in and are values ​​of different types of data, is the mean of the data, W is the weight matrix, and n is the number of data points;

[0093] The temporal autocorrelation calculation formula is: ,in is the length of the time series, is the lag time, is the time series of input data, e is the mean of the time series;

[0094] Step 7: Create the , generated in step 5 , and Tm_WRF_M are input together into a joint model consisting of a deconvolution model and a temporal super-resolution model. The spatial resolution of the output model is set to 1 km, and the temporal resolution is set to 0.5 hours. By running the model, a high spatial resolution feature vector of atmospheric temperature can be generated. , high temporal resolution feature vector , and Tm_WRF_H with high temporal and spatial resolution; similarly, applying the above operations to atmospheric humidity and atmospheric pressure data can obtain the high spatial resolution feature vectors of atmospheric humidity and atmospheric pressure respectively. , , a high temporal resolution feature vector , , and Qm_WRF_H and Pm_WRF_H with high temporal and spatial resolution; then, , and Tm_WRF_H are input into the atmospheric temperature 3D spatiotemporal convolution model established in step 6. By running the model, atmospheric temperature data with high spatiotemporal resolution can be generated. , And Qm_WRF_H are input into the atmospheric humidity 3D spatiotemporal convolution model established in step 6. Running the model can generate atmospheric humidity data with high spatiotemporal resolution. , Input the atmospheric pressure 3D spatiotemporal convolution model established in step 6 together with Pm_WRF_H, and run the model to generate atmospheric pressure data with high spatiotemporal resolution;

[0095] The temporal super-resolution model is: ,in for The time characteristic vector of the atmospheric temperature data at the next moment, is the ratio of high time resolution to original time resolution;

[0096] Step 8: Figure 8 As shown, the two calculation models of atmospheric refractive index and electromagnetic wave propagation path are connected in series, and then the high temporal and spatial resolution atmospheric temperature, humidity and air pressure data obtained in step 7 and the weather radar geographic location information and antenna operation parameter information collected in step 1 are input into the series model together. By running the model, accurate inversion of high temporal and spatial resolution weather radar beam propagation path can be achieved.

[0097] The above is only a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used for various other combinations, modifications and improvements, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art do not deviate from the spirit and scope of the present invention, and should be within the scope of protection of the claims attached to the present invention.

Claims

1. A high temporal and spatial resolution weather radar beam propagation path inversion method, characterized by: The method comprises: S1. Input the collected data into geographic information processing software, calculate the data set A according to the collected data, calculate the spatial distance difference, spatial distance weight value, temporal distance difference and temporal distance weight value according to the data set A, and construct the spatiotemporal distance weight, and use the minimum value of the spatiotemporal distance weight as the screening condition to screen the data set A, and obtain the WRF atmospheric temperature data, WRF atmospheric humidity data and WRF atmospheric pressure data that match the spatiotemporal data; S2, the data generated by S1 are used as input data set 1, the high-frequency radio sounding meteorological data collected by S1 are used as output data set, a multi-scale convolutional neural network is constructed based on the original model of convolutional neural network, and finally the local spatial feature weight coefficients of atmospheric temperature, atmospheric humidity and atmospheric pressure are obtained, a dynamic weight generation model is constructed based on a small fully connected neural network model, and finally the spatial feature vector after weighted fusion of atmospheric temperature, atmospheric humidity and atmospheric pressure is obtained, and a neural network constructed by C-layer 1D convolutional network and a bidirectional LSTM model are constructed, and finally the time feature vector of atmospheric temperature, atmospheric humidity and atmospheric pressure data is obtained; S3, the spatial feature vector and time feature vector after weighted fusion of meteorological data generated by S2, and the data generated by S1 form input data set 2, and calculate its comprehensive complexity, take the high-frequency radiosonde meteorological data collected by S1 as the output data set, introduce the 3D spatiotemporal convolutional network model to construct a 3D spatiotemporal convolutional model of atmospheric temperature, atmospheric humidity and atmospheric pressure; S4, input the spatial feature vector and time feature vector after weighted fusion of the meteorological data generated by S2, and the data generated by S1 into the joint model, and the generated data are respectively input into the 3D spatiotemporal convolution model of atmospheric temperature, atmospheric humidity and atmospheric pressure generated by S3 to obtain atmospheric temperature data, atmospheric humidity data and atmospheric pressure data with high spatiotemporal resolution; S5. Connect the two calculation models of atmospheric refractive index and electromagnetic wave propagation path in series, input the data obtained by S4 and the weather radar geographic location information and antenna operation parameter information data collected by S1 into the series model, and realize the accurate inversion of weather radar beam propagation path with high temporal and spatial resolution.

2. The high temporal and spatial resolution weather radar beam propagation path inversion method according to claim 1, characterized in that: The data collected in S1 include: Collect meteorological data and auxiliary data output by mesoscale WRF at multiple times. The meteorological data output by WRF includes atmospheric temperature, atmospheric humidity and atmospheric pressure. The auxiliary data output by WRF includes the latitude, longitude and altitude of the meteorological data, temporal resolution, spatial resolution and the time corresponding to the data. WRF is a weather forecast mode. Collect high-frequency radiosonde meteorological data and auxiliary data at multiple times. High-frequency radiosonde meteorological data include atmospheric temperature, atmospheric humidity and atmospheric pressure. High-frequency radiosonde auxiliary data include the latitude, longitude and altitude of the data, time resolution, spatial resolution and the time corresponding to the data. Collect weather radar geographic location information and antenna operating parameter information, including the longitude, latitude, altitude, antenna elevation angle and antenna azimuth angle of the weather radar site.

3. The high temporal and spatial resolution weather radar beam propagation path inversion method according to claim 2, characterized in that: The S1 specifically includes the following contents: A1. Input the collected atmospheric temperature data and auxiliary data output by WRF at multiple times into the geographic information processing software, set the spatial interval of the spatial resolution grid to ΔS, select the bilinear interpolation scheme, and run the geographic information processing software to generate WRF atmospheric temperature data with medium spatial resolution, recorded as T_WRF_M, and its time information is recorded as , the spatial position consists of longitude, latitude and altitude, recorded as ; A2. The atmospheric temperature data from the high-frequency radiosonde meteorological data collected at multiple times is recorded as T_Radio, and its time information is recorded as , the spatial position consists of longitude, latitude and altitude, recorded as ; A3. Set the height difference threshold of the adjacent spaces at the same altitude level to , calculate the absolute height difference between a data point P in T_Radio and all T_WRF_M data points , select the height difference less than The corresponding T_WRF_M data points are recorded as data set A; A4. Calculate the spatial distance difference between point P and the data points in data set A according to the spatial distance difference calculation formula , calculate the spatial distance weight value from the data point in data set A to point P according to the spatial distance weight calculation formula ; A5. Calculate the time distance difference between point P and the data point in data set A based on the time data of T_WRF_M and T_Radio. , calculate the time distance weight value from the data point in data set A to point P according to the time distance weight calculation formula ; A6. Constructing space-time distance weights ,by The minimum value is used as the screening condition, and the data points in the screened data set A are used as the spatiotemporal matching points of point P; A7. Repeat steps A3-A6 to process all data points of T_Radio one by one to obtain the WRF atmospheric temperature dataset that matches the T_Radio data in time and space, recorded as Tm_WRF_M. A8. Apply methods A2-A7 to the collected WRF atmospheric humidity and atmospheric pressure data to obtain WRF atmospheric humidity data Qm_WRF_M and atmospheric pressure data Pm_WRF_M that are temporally and spatially matched with the high-frequency radiosonde meteorological data.

4. The high temporal and spatial resolution weather radar beam propagation path inversion method according to claim 3 is characterized by: The S2 specifically includes the following contents: B1. The generated Tm_WRF_M and the corresponding altitude, longitude, latitude, and time data are used as the first input data set. The atmospheric temperature data T_Radio in the collected high-frequency radiosonde meteorological data is used as the output data set. A multi-scale convolutional neural network is constructed based on the original convolutional neural network model. The local spatial feature weight coefficients of atmospheric temperature, atmospheric humidity, and atmospheric pressure are obtained through the multi-scale convolutional neural network. B2. The generated Tm_WRF_M and the corresponding altitude, longitude, latitude and time data are used as the second input data set. The atmospheric temperature data T_Radio in the collected high-frequency radiosonde meteorological data is used as the output data set. A dynamic weight generation model is constructed based on a small fully connected neural network model. The weighted spatial characteristic vector of atmospheric temperature, atmospheric humidity and atmospheric pressure is obtained according to the dynamic weight generation model. B3. The collected data is processed by S1 to obtain Tm_WRF_M and T_Radio at multiple times and the corresponding altitude, longitude, latitude, and time data to form the third input data set. T_Radio at multiple times is used as the output data set. A neural network constructed by a C-layer 1D convolutional network and a bidirectional LSTM model are constructed to generate the time feature vectors of atmospheric temperature, atmospheric humidity, and atmospheric pressure data.

5. The high temporal and spatial resolution weather radar beam propagation path inversion method according to claim 4, characterized in that: The B1 specifically includes the following contents: B11, the generated Tm_WRF_M and the corresponding altitude, longitude, latitude and time data are used as the first input data set, and the atmospheric temperature data T_Radio in the high-frequency radiosonde meteorological data collected at multiple times are used as the output data set; B12. Based on the original model of convolutional neural network, K convolution kernels of different sizes are set, and the dynamic tuning strategy based on evolution is selected to construct a multi-scale convolutional neural network. Each scale of convolution kernel constitutes a convolution channel separately, and convolution operations are performed on the input data set respectively to extract the spatial features of different scales of the input data set, which is recorded as ,in Respectively represent the longitude, latitude, and altitude of the data point, k=1,2,3,…,K; B13. Calculate based on the atmospheric temperature gradient change rate formula and the mean square error formula The gradient change rate and mean square error , the normalized gradient change rate and mean square error are processed using the normalization formula to obtain the normalized gradient change rate and mean square error ; B14. Calculate the local spatial characteristic weight coefficient of atmospheric temperature using convolution kernels of different sizes according to the local weight formula ; B15. Apply the methods of B11-B14 to the generated Qm_WRF_M data and Pm_WRF_M data to obtain the local spatial characteristic weight coefficients of atmospheric humidity and atmospheric pressure respectively. and .

6. The high temporal and spatial resolution weather radar beam propagation path inversion method according to claim 4, characterized in that: The B2 specifically includes the following contents: B21, the generated Tm_WRF_M and the corresponding altitude, longitude, latitude, and time data are used as the second input data set, and the atmospheric temperature data T_Radio in the high-frequency radiosonde meteorological data collected at multiple times are used as the output data set; B22. Based on a small fully connected neural network model, K convolution kernels of different sizes are set. The global average feature extraction formula is used to process each type of data in the input data set to obtain the global feature ; B23. Random sampling is performed in the Gaussian distribution sequence to generate the initial matrix of the connection layer weights, the loss function is set to the mean square error between the input and the input data set, the gradient descent strategy is selected to iteratively optimize the model performance, and a dynamic weight generation model for global features is constructed. The model is run to obtain the global feature weight coefficients of convolution kernels of different scales. ; B24. According to the local spatial feature weight coefficients of convolution kernels of different sizes, a local and global feature weight coefficient balance model is constructed, and the global feature weight coefficient is corrected to obtain the corrected weight coefficient. ; B25, according to Spatial features extracted by convolution of K different scales Perform weighted fusion processing to obtain the spatial characteristics of atmospheric temperature after weighted fusion ; B26. Apply the methods of B21-B25 to the generated Qm_WRF_M data and Pm_WRF_M data to obtain the weighted fusion spatial feature vectors of atmospheric humidity and atmospheric pressure, respectively. and .

7. The high temporal and spatial resolution weather radar beam propagation path inversion method according to claim 6, characterized in that: The B3 specifically includes the following contents: B31, after processing the collected WRF meteorological data and auxiliary data at multiple times, as well as the high-frequency radiosonde meteorological data and auxiliary data, obtain Tm_WRF_M, T_Radio and corresponding altitude, longitude, latitude, and time data at multiple times, use Tm_WRF_M at multiple times and the corresponding altitude, longitude, latitude, and time data as the third input data set, and use T_Radio at multiple times as the output data set; B32. Construct a neural network consisting of C layers of 1D convolutional networks. Each layer of the convolutional network uses a different small-scale convolution kernel to perform convolution operations on the input data set and extract C types of short-time scale change features, recorded as , ,…, ; B33. Set L sliding windows with longer time scales, and record the window sizes as Long_1, Long_2, …, Long_L respectively. Build a bidirectional LSTM model, set the number of LSTM units as NUM, and run the bidirectional LSTM model under different sliding window length conditions to obtain L long time scale change characteristics, recorded as , ,…, ; B34. Combine the short-time scale variation characteristics with the long-time scale variation characteristics to obtain the time characteristic vector of the atmospheric temperature data in the third input data set. ; B35. Apply the methods of B31-B34 to the Qm_WRF_M data and Pm_WRF_M data at multiple times to obtain the time characteristic vectors of atmospheric humidity and atmospheric pressure data respectively. and .

8. The high temporal and spatial resolution weather radar beam propagation path inversion method according to claim 7, characterized in that: The S3 specifically includes the following contents: C1, generate , , and the extracted Tm_WRF_M together constitute the input data set 2. The spatial complexity of the input data set is calculated according to the spatial autocorrelation calculation formula. , the time complexity of the input data set is calculated according to the time autocorrelation calculation formula , add the time complexity and space complexity to get the comprehensive complexity ; C2, the atmospheric temperature data from the high-frequency radiosonde data collected at multiple times are used as the output data set, and the 3D spatiotemporal convolutional network model is introduced. To select the convolution kernel, train the network model, and build a 3D spatiotemporal convolution model of atmospheric temperature; C3. Apply the methods of C1 and C2 to the atmospheric humidity and atmospheric pressure data to obtain the 3D convolution models of atmospheric humidity and atmospheric pressure, respectively.

9. The high temporal and spatial resolution weather radar beam propagation path inversion method according to claim 7, characterized in that: The S4 specifically includes the following contents: D1, generate , , and the extracted Tm_WRF_M are input into a joint model consisting of a deconvolution model and a temporal super-resolution model, and the joint model is run to generate a high spatial resolution feature vector of atmospheric temperature. , high temporal resolution feature vector , and Tm_WRF_H with high temporal and spatial resolution; D2. Apply the method of D1 to the atmospheric humidity and atmospheric pressure data to obtain the high spatial resolution feature vectors of atmospheric humidity and atmospheric pressure respectively. , , a high temporal resolution feature vector , , and Qm_WRF_H and Pm_WRF_H with high temporal and spatial resolution; D3. , The data are input together with Tm_WRF_H into the atmospheric temperature 3D spatiotemporal convolution model, and the model is run to generate atmospheric temperature data with high spatiotemporal resolution; D4. , The data are input into the atmospheric humidity 3D spatiotemporal convolution model together with Qm_WRF_H, and the model is run to generate atmospheric humidity data with high spatiotemporal resolution; D5. , The data are input together with Pm_WRF_H into the 3D spatiotemporal convolution model of atmospheric pressure, and the model is run to generate atmospheric pressure data with high spatiotemporal resolution.

Images