Precipitation model processing method and system based on dual-polarization radar data
By performing hierarchical processing and screening of correlation factors on dual-polarization radar data, and combining Transformer and Kalman filter correction, the problems of slow training speed and low accuracy of existing models are solved, and efficient and reliable precipitation prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HUAZHONG UNIV OF SCI & TECH
- Filing Date
- 2023-11-07
- Publication Date
- 2026-06-23
AI Technical Summary
Existing deep learning-based short-term forecast models are slow to train when processing massive amounts of radar grid data, have limited predictive capabilities, and struggle to achieve high-precision extreme weather precipitation forecasts.
A dual-polarization radar data hierarchical processing method was adopted, dividing the horizontal reflectivity factor ZH into five layers. The mutual information method was used to select factors with high correlation to precipitation as input to the Transformer model. Kalman filtering and AR model were combined to correct forecast errors and improve forecast accuracy.
This significantly improves the accuracy and reliability of precipitation forecasting, reduces model training efficiency, and ensures the practical value of the forecast.
Smart Images
Figure CN117492005B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of hydrology and water resources technology, and in particular to a precipitation model processing method and system based on dual-polarization radar data. Background Technology
[0002] In recent years, with global climate change, natural variations, and changes in terrestrial habitats, a series of extreme and abnormal weather events have occurred frequently, exhibiting significant multi-scale spatiotemporal variability. The frequency and intensity of meteorological disasters such as typhoons, hail, tornadoes, and short-duration heavy rainfall have also shown a year-on-year increasing trend. Among these, extreme severe convective precipitation is a typical type of extreme and hazardous weather. This type of heavy rainfall is highly uneven in both time and space. Therefore, studying the characteristics of heavy rainfall and developing high-precision, high-resolution, real-time quantitative precipitation forecasts to prevent floods is an urgent problem for meteorological departments.
[0003] Currently, precipitation forecasting using dual-polarization radar has become one of the mainstream methods used by meteorological departments. Dual-polarization radar can simultaneously measure radar echoes with both horizontal polarization (H-polarization) and vertical polarization (V-polarization), providing more information than single-polarization radar, including more external characteristics of precipitation particles and more accurate classification information. Because dual-polarization radar can provide reflectivity data for both horizontal and vertical polarization, it can estimate the size and shape of precipitation particles, which plays a significant role in the prediction and monitoring of extreme weather events.
[0004] Traditional short-term forecasting of severe convective weather mainly relies on radar and other observational data, combined with storm identification and tracking technologies for radar extrapolation forecasting. This involves extrapolating the radar reflectivity factor for future times and then using the empirical relationship between the radar reflectivity factor and precipitation (i.e., the ZR relationship) to estimate future precipitation. In recent years, with the accumulation of big data and the development of computing power, artificial intelligence and deep learning technologies have developed rapidly. Deep learning methods are data-driven approaches; theoretically, their performance improves with the amount of training data, making them well-suited for short-term forecasting with large amounts of radar observation data. Currently, there are two main types of deep learning-based short-term forecasting models internationally: one based on convolutional neural networks (CNNs), such as U-Net, and recurrent neural networks (RNNs), such as ConvLSTM and DGMR. However, these models suffer from slow training speeds, limited predictive performance, and poor practical application when faced with massive radar raster datasets, exhibiting generally low accuracy for large-scale depictions. Summary of the Invention
[0005] The purpose of this invention is to disclose a precipitation model processing method and system based on dual-polarization radar data, so as to improve the computational efficiency of the model and ensure the reliability of precipitation assessment results.
[0006] To achieve the above objectives, the method of the present invention includes:
[0007] Step S1: Obtain near-surface precipitation data corresponding to each grid cell within the target analysis area, and horizontal reflectivity factors Z corresponding to at least two different altitudes for dual-polarization radar. H and differential reflectivity Z DR ;
[0008] Step S2: Adjust the horizontal reflectivity factor Z H The target interval is divided into five layers, and each grid cell is arranged according to its Z-axis at the closest height to the ground. H The values are stratified; then, the horizontal reflectivity factor Z corresponding to each grid cell at each height and time period is selected from the candidate factor set. H and differential reflectivity Z DR The values are converted to the average of the summation of non-zero values, and the surface precipitation data is used as the label for the corresponding raster cell.
[0009] Step S3: From the candidate factor set after transformation of each layer, the data sequences corresponding to some factors of each layer that meet the set conditions for correlation with precipitation are selected by mutual information method and used as input to the Transformer forecast model;
[0010] Step S4: After obtaining the precipitation data predicted by the Transformer model, the actual precipitation value is subtracted from the predicted value to obtain the precipitation residual sequence. The Kalman filter method is used to perform noise reduction and smoothing operations on the residual sequence. After performing forecast error cascade correction, the corrected precipitation prediction value is obtained.
[0011] Step S5: Reverse the stratification principle of the predicted precipitation values of each layer to obtain the overall predicted value.
[0012] Preferably, the present invention uses an AR model for cascaded correction of forecast errors, and the prediction residuals are as follows: e t =(e1, e2, … e n In the formula, e t The residual sequence is obtained by subtracting the actual and predicted rainfall values.
[0013] The predicted residuals were used to make AR model predictions. The order of the AR model was determined using the AIC criterion. The residuals obtained from the AR model predictions are as follows: In the formula, These are the predicted values obtained from the AR model;
[0014] The final revised precipitation forecast value was obtained. for:
[0015] ;
[0016] Among them, R t These are the precipitation predictions output by the Transformer model.
[0017] Preferably, in the process of screening out the factors of each layer that satisfy the set conditions for correlation with precipitation using the mutual information method, the specific formula for calculating the correlation I(Z;R) is as follows:
[0018] ;
[0019] In the formula, p(x, y) is the joint density function of radar reflectivity Z and precipitation R, and p(x) and p(y) are the marginal density functions of Z and R, respectively.
[0020] To achieve the above objectives, the present invention also discloses a precipitation model processing system based on dual-polarization radar data, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-described method when executing the computer program.
[0021] This invention has a clear concept, is easy to operate, and is highly practical; moreover, it has the following beneficial effects:
[0022] The horizontal reflectivity factor Z H Perform layered processing, using Z H The layer values represent rainfall intensity, transforming complex raster data into time-series numerical sequences. Simultaneously, using mutual information, data sequences corresponding to factors in each layer that satisfy set conditions for correlation with precipitation are independently selected as input to the Transformer forecast model, further achieving data dimensionality reduction and significantly reducing the training efficiency of traditional models. Furthermore, after obtaining the precipitation data predicted by the Transformer model, Kalman filtering is used to perform cascade correction on the prediction results, ensuring the accuracy and reliability of precipitation prediction. This approach has strong practical value and can significantly improve forecast accuracy and reliability compared to commonly used radar wave convolution prediction methods.
[0023] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description
[0024] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:
[0025] Figure 1 This is a schematic diagram of the precipitation model processing method based on dual-polarization radar data disclosed in an embodiment of the present invention. Detailed Implementation
[0026] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.
[0027] Example 1
[0028] This embodiment discloses a precipitation model processing method based on dual-polarization radar data, which specifically includes the following steps:
[0029] Step 1: Collect, organize, and verify the horizontal reflectivity factor Z at different altitudes of the dual-polarization radar. H Differential reflectivity Z DR And near-surface precipitation data (no altitude).
[0030] Dual-polarization radar is a radar system that uses two different polarizations, typically horizontal (H) and vertical (V). It transmits horizontally or vertically polarized microwave pulse signals via an antenna system, and simultaneously receives signals from both polarizations. The transmitted microwave signals propagate through the atmosphere and encounter various targets, resulting in reflection, scattering, and absorption, such as precipitation, particulate matter, and vegetation. The received dual-polarization signals, after data processing and algorithm analysis, can measure the reflection of electromagnetic waves from precipitation particles in both the horizontal and vertical directions, thereby obtaining microphysical information such as the size, phase, and water content of the precipitation particles. Furthermore, the radar system can scan different altitude layers by adjusting the antenna elevation angle. At different elevation angles, the radar system can measure reflected signals at different altitudes. Typically, the radar system scans a series of elevation angles, covering the vertical altitude range of interest. The two most commonly used variables in dual-polarization radar precipitation assessment are: 1) Z H 1) Horizontal reflectivity factor, which is the echo intensity in the horizontal direction, usually measured in dBZ, mainly reflects the intensity of precipitation; 2) Z DR Differential reflectivity, which is the difference in echo intensity in the horizontal and vertical directions, mainly reflects the size of precipitation particles in the observation area. Near-surface precipitation data is acquired by meteorological satellites and then processed to correspond to the grid points of radar reflectivity data.
[0031] Redundant data (radar station, building reflectance values) and outliers are removed; values exceeding the range are replaced with the mean of surrounding values, where Z... H Value range: [0, 65], Z DR Value range: [-1, 5].
[0032] Step 2: Since there is an empirical relationship between radar reflectivity and precipitation, it is usually expressed as the following formula: R = aZ b ;
[0033] In the formula, R is precipitation, Z is radar reflectivity, and a and b are empirical parameters that usually vary in different regions and precipitation types.
[0034] There is an exponential relationship between radar reflectivity and precipitation intensity; the higher the reflectivity, the stronger the rainfall. Therefore, this invention uses the horizontal reflectivity factor Z at different altitudes. H The target interval (0, 65] is divided into five layers. This equal division is to reduce the data dimensionality, reducing the original 256*256 (corresponding to the target analysis area, equivalent to the area of some provinces in China, in km²) to five layers. 2 The raster data was reduced to 5 data sequences. Simultaneously, according to Z... H The resulting grid (i.e., grid cell, typically 1*1 in size, measured in km) 2 Location mapping to Z DR From the near-surface precipitation data R, Z at different altitudes can be obtained. DR Five-layer data and five-layer near-surface precipitation data. When dividing the layers, the average of the non-zero values in each layer is taken as the value of that layer (e.g., Z). H The data for the five grid cells in the first layer (0, 13) are: 3.1, 3, 2.1, 4.2, 5.4. If these five numbers correspond to the Z-axis at a height of 1 km... DR The possible values are 43, 3.2, 5.3, 0.0, and 0.5. Calculate Z. DR The mean is not accumulated by 0.0; instead, the average of the four non-zero values is calculated and then mapped to the corresponding five raster cells to obtain the Z-value. DR (Values), different levels correspond to different intensities of precipitation, with the precipitation intensity increasing sequentially from level 1 to level 5.
[0035] In this embodiment, the mapping process in the above-mentioned layering process is: to map the horizontal reflectivity factor Z... H The target interval is divided into five layers, and each grid cell is arranged according to its Z-axis height closest to the ground. H The values are stratified, and then the horizontal reflectivity factor Z corresponding to each grid cell at each height and time period is selected from the candidate factor set. H and differential reflectivity Z DR The values are converted to the average of the sum of non-zero values. Correspondingly, the surface precipitation data serves as the label for the corresponding raster cell.
[0036] In a specific application example, each grid in this step represents a frame of data obtained after a radar scan, with a 6-minute time interval between adjacent frames. Near-surface precipitation data is acquired through meteorological satellites, processed, and then correlated with radar data grid points to represent the cumulative precipitation value over a 6-minute period.
[0037] Step 3: Divide the layered frame-by-frame Z-sectionH and Z DR The set of value factors serves as the candidate set of factors for precipitation data Rt.
[0038] , g = 1,3,7 i = 1,2,3,4,5;
[0039] In the formula, It is the radar scanning altitude. The values of 1, 3, and 7 represent the ground clearance of 1 km, 3 km, and 7 km based on parameters measured at different elevation angles using a dual-polarization radar. Indicates the first Layer data, where t represents the time period, i.e., the number of frames. Taking layer 1 as an example, Z at different heights... H and Z DR The values of 11 time periods are taken as the first layer of precipitation R. t The candidate factor set has three levels and 66 factor sets.
[0040] Step 4: After dividing each frame of radar reflectance and precipitation data into 5 layers, factors with high correlation to precipitation need to be selected for use as input to the forecast model. This embodiment uses mutual information (MI) to evaluate the Z values for layers 1-5 respectively. H and Z DR The correlation between the value factor set and precipitation data Rt shows that mutual information can not only capture linear relationships but also detect nonlinear relationships between variables, making it effective in studying correlations in complex data.
[0041] ;
[0042] In the formula, p(x, y) is the joint density function of radar reflectivity Z and precipitation R, and p(x) and p(y) are the marginal density functions of Z and R, respectively.
[0043] The greater the correlation, the stronger the interdependence between the factor and precipitation. When screening factors for forecast models, factors with high correlation should be given priority. This can not only reduce the dimensionality of the model input and reduce redundant information, but also reduce the risk of overfitting.
[0044] Each layer selects the top 30 input factors with the highest mutual information (MI) as the driving factors for the attention mechanism Transformer precipitation estimation model, and divides the training set and test set data in an 8:2 ratio, predicting precipitation Rt for layers 1 to 5 respectively. This ensures that the features of the data input sequences corresponding to each layer used for subsequent precipitation model predictions are typically inconsistent, guaranteeing the accuracy of data processing at each layer. Furthermore, in a specific experimental scenario of this embodiment, when processing 34,987 frames of raster data (one frame representing one radar scan) continuously scanned from the same area by the same radar, the aforementioned correlation feature factor selection only requires one round of selection, meaning that only five layers of correlation feature selection are needed.
[0045] Step 4: Make predictions based on the Transformer forecasting model.
[0046] In this step, the Transformer model completely abandons the structures of RNNs and CNNs, instead employing a decode-encoder model composed of a self-attention mechanism and a feedforward neural network. Because it does not require a recurrent architecture for sequence alignment, it can be trained in parallel much faster. The Transformer model mainly consists of an input layer, an encoder layer, a decoder layer, and an output layer. Initially widely used in natural language processing and image recognition, it has gradually gained attention in the field of time series prediction.
[0047] Step 5: After obtaining the precipitation data predicted by Transformer, the actual precipitation value is subtracted from the predicted value to obtain the precipitation residual sequence. Kalman filtering is used to smooth and denoise the residual sequence. An AR model is then used for cascaded correction of forecast errors. The prediction residual is as follows: e t =(e1, e2, … e n In the formula, This is the residual sequence obtained by subtracting the actual rainfall value from the predicted rainfall value.
[0048] The predicted residuals were used to make AR model predictions. The order of the AR model was determined using the AIC criterion. The residuals obtained from the AR model predictions are as follows: In the formula, These are the predicted values obtained from the AR model.
[0049] The final corrected precipitation forecast is as follows:
[0050] ;
[0051] In the formula, R t These are the precipitation predictions output by the Transformer model.
[0052] Step 6: By reverse-engineering the predicted precipitation values for layers 1 to 5 according to the above stratification approach, the overall predicted value can be obtained.
[0053] In summary, referring to Figure 1 Those skilled in the art will conclude that the core of this embodiment is the following steps:
[0054] Step S1: Obtain near-surface precipitation data corresponding to each grid cell within the target analysis area, as well as the horizontal reflectivity factor and differential reflectivity at at least two different altitudes corresponding to the dual-polarization radar.
[0055] Step S2: Divide the horizontal reflectivity factor into five layers according to the target interval, and then divide each grid cell into layers according to the value of the horizontal reflectivity factor at the height closest to the ground. Then, in the candidate factor set, convert the horizontal reflectivity factor and differential reflectivity of the corresponding grid cells at each height and time period into the average value after accumulating non-zero values, and use the ground precipitation data as the label of the corresponding grid cell.
[0056] Step S3: From the candidate factor set after transformation of each layer, the data sequences corresponding to some factors of each layer that meet the set conditions for correlation with precipitation are selected by mutual information method and used as input to the Transformer forecast model.
[0057] Step S4: After obtaining the precipitation data predicted by the Transformer model, the actual precipitation value is subtracted from the predicted value to obtain the precipitation residual sequence. The Kalman filter method is used to perform noise reduction and smoothing operations on the residual sequence. After performing forecast error cascade correction, the corrected precipitation prediction value is obtained.
[0058] Step S5: Reverse the stratification principle of the predicted precipitation values of each layer to obtain the overall predicted value.
[0059] Example 2
[0060] Corresponding to the above method, this embodiment discloses a precipitation model processing system based on dual-polarization radar data, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the series of steps in the above method embodiment.
[0061] In summary, the method and system disclosed in this invention are clear in concept, easy to operate, and highly practical; moreover, they have the following beneficial effects:
[0062] The horizontal reflectivity factor Z H Perform layered processing, using Z HThe layer values represent rainfall intensity, transforming complex raster data into time-series numerical sequences. Simultaneously, using mutual information, data sequences corresponding to factors in each layer that satisfy set conditions for correlation with precipitation are independently selected as input to the Transformer forecast model, further achieving data dimensionality reduction and significantly reducing the training efficiency of traditional models. Furthermore, after obtaining the precipitation data predicted by the Transformer model, Kalman filtering is used to perform cascade correction on the prediction results, ensuring the accuracy and reliability of precipitation prediction. This approach has strong practical value and can significantly improve forecast accuracy and reliability compared to commonly used radar wave convolution prediction methods.
[0063] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A precipitation model processing method based on dual-polarization radar data, characterized in that, include: Step S1: Obtain near-surface precipitation data corresponding to each grid cell within the target analysis area, and horizontal reflectivity factors Z corresponding to at least two different altitudes for dual-polarization radar. H and differential reflectivity Z DR ; Step S2: Adjust the horizontal reflectivity factor Z H The target interval is divided into five layers, and each grid cell is arranged according to its Z-axis height closest to the ground. H The values are stratified, and then the horizontal reflectivity factor Z corresponding to each grid cell at each height and time period is selected from the candidate factor set. H and differential reflectivity Z DR The values are converted to the average of the summation of non-zero values, and the surface precipitation data is used as the label for the corresponding raster cell. Step S3: From the candidate factor set after transformation of each layer, the data sequences corresponding to some factors of each layer that meet the set conditions for correlation with precipitation are selected by mutual information method and used as input to the Transformer forecast model; Step S4: After obtaining the precipitation data predicted by the Transformer model, the actual precipitation value is subtracted from the predicted value to obtain the precipitation residual sequence. The Kalman filter method is used to perform noise reduction and smoothing operations on the residual sequence. After performing forecast error cascade correction, the corrected precipitation prediction value is obtained. Step S5: Reverse the stratification principle of the predicted precipitation values of each layer to obtain the overall predicted value.
2. The method according to claim 1, characterized in that, The prediction residuals are obtained by using an AR model for cascaded correction of prediction errors: the t =(e1,e2,...,e n ); In the formula, e t The residual sequence is obtained by subtracting the actual and predicted rainfall values. The predicted residuals were used to make AR model predictions. The order of the AR model was determined using the AIC criterion. The residuals obtained from the AR model predictions are as follows: In the formula, These are the predicted values obtained from the AR model; The final revised precipitation forecast value was obtained. for: Among them, R t These are the precipitation predictions output by the Transformer model.
3. The method according to claim 1 or 2, characterized in that, In the process of using the mutual information method to screen out the factors in each layer that meet the set conditions for correlation with precipitation, the specific formula for calculating the correlation I(Z;R) is as follows: In the formula, p(x,y) is the joint density function of radar reflectivity Z and precipitation R, and p(x) and p(y) are the marginal density functions of Z and R, respectively.
4. A precipitation model processing system based on dual-polarization radar data, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method described in any one of claims 1 to 3.