Nowcasting method of radar extrapolation and ZR relationship based on evolutionary network and Unet network

Abstract

The present invention discloses a nowcast precipitation forecasting method based on radar extrapolation and the Z-R relationship using an evolutionary network and a Unet network. The method comprises the following steps: obtaining historical actual rainfall data for hourly periods and using it to construct a radar echo sample dataset; constructing a radar echo short-term nowcasting prediction model, comprising an evolutionary network module and a Unet variant module; constructing a model training loss function, and training the radar echo short-term nowcasting prediction model in conjunction with the radar echo sample dataset; obtaining a radar mosaic echo one hour before the prediction time, inputting it into the trained model to obtain a predicted radar echo, and obtaining a precipitation grid forecast field based on the Z-R relationship and the predicted radar echo. By designing the evolutionary network module and the Unet variant module, as well as loss functions for the generator and the discriminator, the present invention improves the forecast accuracy of near-term heavy precipitation, has good application prospects, and provides more accurate forecasting services for disaster prevention and mitigation.

Description

Technical Field

[0001] The present invention belongs to the technical field of weather forecasting, and in particular relates to a nowcasting method based on radar extrapolation and ZR relationship of an evolutionary network and a Unet network. Background Art

[0002] Precipitation forecasting, a crucial element of nowcasting, is crucial for modern urban management and directly impacts land and air transportation management, large-scale event support, and flood control. High-precision nowcasting is essential for future smart city development. However, the forecast information currently available to forecasters is insufficiently targeted. The TS scores for objective forecasts of near-term heavy precipitation fall significantly short of the actual forecast requirements for disaster prevention and mitigation of major meteorological disasters, impacting disaster prevention and mitigation decision-making. Meteorological authorities urgently need to improve the accuracy of weather forecasts.

[0003] Nowcasting based on dense observational data, such as weather radar, is a primary tool for heavy precipitation monitoring and warning. Radar nowcasting is essentially a spatiotemporal sequence prediction task, which can also be viewed as a video prediction task. Using past radar echo patterns as input to predict future radar echo patterns, traditional extrapolation methods, based on the assumption of linear evolution of radar echoes, cannot capture nonlinear processes such as the formation and dissolution of convective systems, resulting in significant limitations in practical applications. Recent deep learning techniques have been shown to better account for the spatiotemporal variations and inherent correlations of convective systems. Therefore, applying deep learning to address precipitation nowcasting is a promising approach and a new approach to improving heavy precipitation nowcasting. Currently, refined grid-based precipitation nowcasting based on radar extrapolation suffers from increased cumulative error, ultimately leading to increased forecast error. With longer forecast horizons, forecast results become "fuzzy," and current nowcasting skill scores are low. Summary of the Invention

[0004] To solve the above technical problems, the present invention proposes a nowcasting method for precipitation based on radar extrapolation and ZR relationship of evolutionary network and Unet network to solve the problems existing in the above-mentioned prior art.

[0005] To achieve the above objectives, the present invention provides a nowcasting method for precipitation based on radar extrapolation and ZR relationship of evolutionary networks and Unet networks, comprising:

[0006] Acquire historical real-time rainfall data for each hour, and construct a radar echo sample data set based on the historical real-time rainfall data;

[0007] Constructing a radar echo short-time nowcasting extrapolation prediction model; the generator of the radar echo short-time nowcasting extrapolation prediction model includes an evolutionary network module and a Unet variant module;

[0008] Constructing a model training loss function, and training the radar echo short-time nowcasting prediction model based on the radar echo sample data set and the model training loss function;

[0009] The radar mosaic echo one hour before the prediction time is obtained, and the trained radar echo short-term nowcasting prediction model is input to obtain the predicted radar echo. The precipitation grid forecast field is obtained based on the ZR relationship and the predicted radar echo.

[0010] Optionally, the real-time rainfall data includes real-time rainfall observation data from ground meteorological stations and real-time observation data from weather radars; the real-time observation data from weather radars is a radar combined reflectivity puzzle.

[0011] Optionally, the process of constructing a radar echo sample dataset includes:

[0012] Each hour within the corresponding time range of the historical real-time rainfall data is used as the starting time for precipitation in the next hour. The radar combined reflectivity puzzle and precipitation are obtained for every 6 minutes one hour before the starting time. The radar combined reflectivity puzzles with the top six digits sorted by the time interval between the acquisition time and the starting time are extracted and overlapped. The segmented grids are filtered according to the proportion of echo intensity. After all the screening is completed, they are combined with the precipitation to obtain the radar echo sample dataset.

[0013] Optionally, the radar echo sample data set is decomposed into two components by the evolutionary network module, and the evolutionary operator is calculated on the two components to obtain preliminary radar echo extrapolation data; the radar echo sample data set and the preliminary radar echo extrapolation data are used as inputs of the Unet variant module.

[0014] Optionally, on the basis of Unet, an attention mechanism, a hole convolution module and a residual module are introduced to obtain the Unet variant module.

[0015] Optionally, the model training loss function includes a weighted loss function and a power spectrum loss function; the generator training of the model adopts the weighted loss function; and the discriminator training of the model adopts the power spectrum loss function.

[0016] Optionally, the weighted loss function is:

[0017]

[0018] Among them, WeightedLoss is the total loss of the weighted loss function; N is the total number of samples; ωi is the weight of the i-th sample, usually used to adjust the loss of the minority class; y i The true label of the i-th sample; The predicted label of the i-th sample.

[0019] Optionally, the power spectrum loss function is:

[0020] L PSD =||P model (f)-P target (f)|| 2

[0021] Among them, L PSD is the power spectral density loss; P model (f) is the power spectrum density of the model output signal; P target (f) is the power spectral density of the target signal, and ||·|| represents the norm.

[0022] Compared with the prior art, the present invention has the following advantages and technical effects:

[0023] The model uses an evolutionary network module to calculate the input radar echo sample dataset, generating a preliminary extrapolated echo dataset. This extrapolated echo dataset and the initial radar echo sample dataset are then fed into the Unet variant module, ultimately yielding the output of the entire generator. This forecast model improves the prediction capability of heavy rainfall exceeding 20 mm by over 15% compared to current near-fall forecasts. This model can effectively improve the accuracy of near-fall forecasts, has promising application prospects, and can provide more accurate forecasting services for disaster prevention and mitigation.

[0024] The present invention designs a power spectrum loss function for the discriminator of the forecast model, which can better characterize the characteristics of radar echo data, more directly evaluate the energy distribution difference between the radar echo data generated by the generator and the real data, and drive the generator to generate a power spectrum distribution that is closer to the real data in the frequency domain. For the generator of the forecast model, a weighted loss function is designed. It is the sum of the weighted losses of the weighted mean absolute error and the weighted mean square error. The weighting coefficient here is assigned different weights according to the actual radar echo intensity. That is, the stronger the radar echo intensity, the greater the weight assigned. This weighting rule increases the information feedback capability of the strong radar echo area so that it will not be averaged out by large areas of weak echo areas. This achieves high-precision forecasts of precipitation echoes above 30dBZ, with higher forecast accuracy in disaster prevention and mitigation warnings. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0026] Figure 1 Flowchart of an objective precipitation nowcasting method based on radar extrapolation and ZR relationship of an evolutionary network module and Unet-Attention-Aspp network structure (EAAUnet) according to an embodiment of the present invention;

[0027] Figure 2 The EAAUnet radar extrapolation technology processing model of the embodiment of the present invention;

[0028] Figure 3 A schematic diagram of actual precipitation according to an embodiment of the present invention;

[0029] Figure 4 Schematic diagram of precipitation forecast according to an embodiment of the present invention. DETAILED DESCRIPTION

[0030] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0031] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0032] Example 1

[0033] like Figure 1-4 As shown, this embodiment provides a nowcasting method for precipitation based on radar extrapolation and ZR relationship of evolutionary networks and Unet networks, including:

[0034] Step S1, obtaining the real-time rainfall observation data from the ground meteorological station and the real-time observation data from the weather radar.

[0035] Specifically, the real-time rainfall observation data of ground meteorological stations are obtained from the TianQing database, which is the real-time rainfall in multiple time periods with 1 hour as the node within the historical time period of each station; the radar real-time observation data are obtained from the SWAN system issued by the China Meteorological Administration, and the radar real-time observation data are radar combined reflectivity puzzle data.

[0036] Step S2: constructing a radar echo extrapolation sample data set.

[0037] Specifically, the data used was the Swan radar mosaic, sampled from July 2016 to June 2022. During this period, for every hour when the number of stations with a density greater than 10 mm reached 20 or more, 30 consecutive frames of radar echoes were extracted every 12 minutes to form a preliminary sample. These preliminary radar echo samples were then segmented with overlap, converting the original 900×1200 grid into 20 256×256 grids. Within these 20 grids, radar echo grid samples with a density greater than 30 dBZ exceeding 15% were selected for inclusion in the final forecast sample dataset, ultimately constructing a total of 33,244 samples.

[0038] Step S3, obtaining preliminary radar echo extrapolation data for the next hour through extrapolation of the evolutionary network module;

[0039] Specifically, to reduce the pressure of model training, the past six consecutive frames of echo are used as model input, and the radar echo map of the next 10 frames, that is, the radar echo map of the next hour, is output. This is achieved here through the evolutionary network module of the extrapolation model. The model input is mainly decomposed into the speed V and intensity S components of the movement through evolutionary coding. These two components are calculated through the operation of the evolutionary operator to obtain preliminary radar echo extrapolation data.

[0040] Step S4, constructing a Unet variant module to obtain the input of the EAAUnet radar extrapolation model discriminator;

[0041] Specifically, the radar data is first input into the evolutionary network module, and then the output of this evolutionary network module is combined with the previous model input as the input of the Unet variant module, and finally the output of the entire generator is obtained.

[0042] Step S5, constructing the EAAUnet radar extrapolation model;

[0043] The main structure of the model is generally an adversarial network structure, with the generator having two parts: an evolutionary network module and a Unet variant module. As the discriminator of this adversarial network, this embodiment uses a simple Unet network. The specific operation process of the model is as follows: the radar data input is first input into the evolutionary network module, and then the output of this evolutionary network module is combined with the previous model input as the input of the Unet variant module, and finally the output of the entire generator is obtained. During model training, these three modules use different loss functions to update the model, including a weighted loss function and a power spectrum (PSD) loss function.

[0044] Step S6, constructing a loss function for model training;

[0045] Specifically, in order to better train the model, this embodiment modifies the loss function. In model training, the loss function is used to evaluate the fit of the current training model and to modify its weight according to the value of the loss function.

[0046] A weighted loss function is designed for the generator, which is the sum of the weighted losses of the weighted mean absolute error and the weighted mean square error. Its general form can be expressed as:

[0047]

[0048] Where: WeightedLoss is the total loss of the weighted loss function; N is the total number of samples; ω i is the weight of the i-th sample, usually used to adjust the loss of the minority class; y i The true label of the i-th sample; The predicted label of the i-th sample.

[0049] The weighting coefficient is assigned different weights according to the actual radar echo strength. That is, the stronger the radar echo strength, the greater the weight assigned. The purpose of this is to increase the information feedback capability of the strong radar echo area so that it is not averaged by the large area of ​​weak echo area. Set the actual radar echo strength to 9 levels:

[0050] Level 1 echo intensity: 0~15dBZ, level 2 echo intensity: 15~25dBZ, level 3 echo intensity: 25~35dBZ, level 4 echo intensity: 35~40dBZ, level 5 echo intensity: 40~45dBZ, level 6 echo intensity: 45~50dBZ, level 7 echo intensity: 50~55dBZ, level 8 echo intensity: 55~65dBZ, level 9 echo intensity: ≥65dBZ.

[0051] A power spectrum loss function (PSDLoss) is designed for the discriminator. It is a loss function used in the fields of signal processing and system identification. In particular, during model training, when the frequency characteristics of the signal need to be optimized, this loss function measures the difference between the power spectrum density (PSD) of the model output signal and the PSD of the target signal. Power spectrum density is the distribution of signal energy in the frequency domain and is usually used to describe the frequency characteristics of the signal. PSDLoss calculates the loss by comparing the PSD of the model output signal with the PSD of the real signal or the expected signal, thereby prompting the model to learn the correct frequency response. Its basic expression is:

[0052] L PSD =||P model (f)-P target(f)|| 2

[0053] Among them, L PSD is the power spectral density loss; P model (f) is the power spectrum density of the model output signal; P target (f) is the power spectral density of the target signal; ||·|| represents some norm, usually the L2 norm (Euclidean norm), but other norms can also be used.

[0054] The use of the power spectrum loss function has three characteristics: first, it can better characterize the characteristics of the echo data; second, it can more directly evaluate the energy distribution difference between the echo data generated by the generator and the real data; third, it drives the generator to generate a power spectrum distribution that is closer to the real echo data in the frequency domain.

[0055] Step S7, training and obtaining a prediction model;

[0056] Step S8: Using the ZR relationship to perform calculation and conversion processing from radar echo to precipitation, and generate a precipitation grid forecast field.

[0057] Specifically, the ZR (Zawadzki-Renard relation) is used to invert precipitation from radar echoes, and the grid point forecast value of precipitation is obtained through calculation. The calculation formula is as follows:

[0058] Z=aR b

[0059] Where Z is the radar reflectivity factor, typically expressed in mm^6 / m^3; R is the precipitation rate, expressed in mm / h; a and b are constants whose values ​​depend on the precipitation type and radar wavelength. Here we set a = 300 and b = 1.4.

[0060] This example primarily constructs an adversarial network deep learning model consisting of an evolutionary network module and a Unet variant module. It employs an overlapping segmentation technique for sample selection and constructs a model input consisting of two distinct sets of radar data. The main network architecture builds on Unet by adding an attention mechanism, a dilated convolution module, and a residual module to enhance the model's feature extraction and extrapolation capabilities. To further optimize model training, the loss function is further adaptively modified. These are the key differences between this method and existing methods for nowcast precipitation prediction based on radar echo extrapolation.

[0061] The following example uses the precipitation forecast for the next hour starting at 08:00 on July 29, 2023 as an example:

[0062] (1) Collect sample data. The statistical period is from June 2016 to June 2022. The samples from June 2016 to December 2021 are used to train the forecast model; the samples from January 2022 to June 2022 are used to test the forecast model. The hourly time of the hour in this period is used as the starting time for precipitation in the next hour. The collected data are: the ground rainfall observation data of the ground meteorological station every hour and the SWAN radar combined reflectivity puzzle with a frame every 6 minutes. The ground rainfall observation data is the precipitation in the past hour.

[0063] (2) Radar sample data processing. The SWAN radar reflectivity mosaic is a 900*1200 size mosaic grid data. The last 6 frames of radar reflectivity mosaics one hour before the start time are cropped into 256×256 grid areas with partial overlap. If the proportion of the 20 256×256 grids with a value greater than 30dBZ reaches more than 15%, this grid is selected into the final data set, totaling 33,244 samples. The data set is divided into: 1) Training sample set: samples from July 2016 to December 2021 (29,000); 2) Test sample set: samples from January 2022 to June 2022 (4,244).

[0064] (3) Construct the input of the forecast model. Use the radar echo matrix (s1_data) of the past 6 consecutive frames in the time period as the input of the ENetwork model, and output the radar echo of the next 10 frames (1 hour). Through evolutionary coding, the input of the model is decomposed into the velocity V and intensity S components of the movement. These two components are calculated by the evolutionary operator to obtain the preliminary echo extrapolation and the preliminary echo extrapolation moment s2_data. The s1_data and s2_data together, a total of 16 channel data, are used to construct the final forecast model input matrix s3_data. At this time, the dimension of the s3_data tensor is (batchsize, 16, 256, 256), where batchsize is the batch size of the training input, 16 is the 16 channel data, and 256 is the width and height of the data.

[0065] (4) Construct the EAAUnet radar extrapolation prediction model. The 16-channel input matrix s3_data is used as the input of the EAAUnet model, which is further trained and the optimal model is saved. During model training, different loss functions are used to train the generator and discriminator of the model:

[0066] (a) Generator loss function design: The generator loss function adopts a weighted loss function. That is, according to the intensity of the radar echo, the following levels are graded: the first-level echo intensity is 0-15dBZ, the second-level echo intensity is 15-25dBZ, the third-level echo intensity is 25-35dBZ, the fourth-level echo intensity is 35-40dBZ, the fifth-level echo intensity is 40-45dBZ, the sixth-level echo intensity is 45-50dBZ, the seventh-level echo intensity is 50-55dBZ, the eighth-level echo intensity is 55-65dBZ, and the ninth-level echo intensity is greater than or equal to 65dBZ. The weights assigned to each level are: first-level echo intensity: 1, second-level echo intensity: 2, third-level echo intensity: 5, fourth-level echo intensity: 10, fifth-level echo intensity: 20, sixth-level echo intensity: 50, seventh-level echo intensity: 60, eighth-level echo intensity: 90, and ninth-level echo intensity: 100.

[0067] The final loss function is:

[0068]

[0069] (b) Discriminator loss function design: The discriminator loss function adopts a power spectrum loss function (PSDLoss), which is a loss function used in the field of signal processing and system identification. In particular, during model training, when the frequency characteristics of the signal need to be optimized, this loss function measures the difference between the power spectrum density (PSD) of the model output signal and the PSD of the target signal. Power spectrum density is the distribution of signal energy in the frequency domain and is usually used to describe the frequency characteristics of the signal. PSDLoss calculates the loss by comparing the PSD of the model output signal with the PSD of the real signal or the expected signal, thereby prompting the model to learn the correct frequency response. Its basic expression is:

[0070] L PSD =||P model (f)-P target (f)|| 2

[0071] Among them, L PSD is the power spectral density loss; P model (f) is the power spectral density P of the model output signal target (f) is the power spectral density of the target signal; ||·|| represents a certain norm, and L2 is the Euclidean norm.

[0072] (5) Calculate the latest forecast sample. For the precipitation forecast for the next hour at 08:00 on July 29, 2023, the starting time is 08:00 on July 29, 2023. The SWAN system is used to obtain the past six radar echo mosaic frames (i.e., six radar echo frames at 07:30 on July 29, 2023, 07:36 on July 29, 2023, 07:42 on July 29, 2023, 07:48 on July 29, 2023, 07:54 on July 29, 2023, and 08:00 on July 29, 2023). Since the SWAN mosaic grid size is 900 × 1200, the model input X matrix shape is (6, 900, 1200), where 6 is the number of the past six consecutive echo frames, and 900 and 1200 are the height and width of the radar echo.

[0073] (6) Extrapolation using the trained model. First, the model input X in (6) is divided into 20 (6, 256, 256) tensors according to the overlapping segmentation method in (2). Each of these is input into the trained model, thereby obtaining the corresponding 20 (10, 256, 256) radar echo extrapolations. The 20 (10, 256, 256) radar echo extrapolations are further reverse-joined, and the average value of the overlapping grid area is calculated. Finally, the radar echo extrapolation of the entire area for the next 10 frames (10, 900, 1200) is obtained. That is, the radar echo for the next hour (each radar echo frame is 6 minutes apart, and 10 frames is 1 hour).

[0074] (7) Radar echo inversion precipitation calculation. The ZR relationship is used to invert the radar echo precipitation, calculate the precipitation amount for the grid point in the next hour, and complete the precipitation forecast for the grid point in the next hour. Figure 3 and Figure 4 The figure is a comparison between the forecast and the actual situation using this method.

[0075] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A nowcasting method based on radar extrapolation and ZR relationship of evolutionary network and Unet network, characterized by: The following steps are involved: Acquire historical real-time rainfall data for each hour, and construct a radar echo sample data set based on the historical real-time rainfall data; Constructing a radar echo short-time nowcasting extrapolation prediction model; the generator of the radar echo short-time nowcasting extrapolation prediction model includes an evolutionary network module and a Unet variant module; Constructing a model training loss function, and training the radar echo short-time nowcasting prediction model based on the radar echo sample data set and the model training loss function; Obtain radar mosaic echoes one hour before the prediction time, input the trained radar echo short-term nowcasting prediction model, obtain predicted radar echoes, and obtain a precipitation grid forecast field based on the ZR relationship and the predicted radar echoes; The radar echo sample data set is decomposed into two components by the evolutionary network module, and the evolutionary operator is calculated on the two components to obtain preliminary radar echo extrapolation data; the radar echo sample data set and the preliminary radar echo extrapolation data are used as inputs of the Unet variant module; On the basis of Unet, the attention mechanism, the dilated convolution module and the residual module are introduced to obtain the Unet variant module; The model training loss function includes a weighted loss function and a power spectrum loss function; the model generator training adopts the weighted loss function; the model discriminator training adopts the power spectrum loss function; The weighted loss function is: Among them, WeightedLoss is the total loss of the weighted loss function; N is the total number of samples; ω i is the weight of the i-th sample, usually used to adjust the loss of the minority class; y i The true label of the i-th sample; The predicted label of the i-th sample; The power spectrum loss function is: L PSD =||P model (f)-P target (f)|| 2 Among them, L PSD is the power spectral density loss; P model (f) is the power spectrum density of the model output signal; P target (f) is the power spectral density of the target signal, and ||·|| represents the norm.

2. The method for nowcasting precipitation based on radar extrapolation and ZR relationship of evolutionary network and Unet network according to claim 1, characterized in that: The real-time rainfall data includes the real-time rainfall observation data of the ground meteorological station and the real-time observation data of the weather radar; the real-time observation data of the weather radar is a radar combined reflectivity puzzle.

3. The method for nowcasting precipitation based on radar extrapolation and ZR relationship of evolutionary network and Unet network according to claim 2, characterized in that: The process of building a radar echo sample dataset includes: Each hour within the corresponding time range of the historical real-time rainfall data is used as the starting time for precipitation in the next hour. The radar combined reflectivity puzzle and precipitation are obtained for every 6 minutes one hour before the starting time. The radar combined reflectivity puzzles with the top six digits sorted by the time interval between the acquisition time and the starting time are extracted and overlapped. The segmented grids are filtered according to the proportion of echo intensity. After all the screening is completed, they are combined with the precipitation to obtain the radar echo sample dataset.

Images