A short-term rainfall forecasting method and system based on raindrop spectrometer and meteorological radar

Abstract

The present invention relates to the field of meteorological and hydrological technology, and discloses a short-term rainfall forecasting method and system based on a raindrop spectrometer and a meteorological radar, the method comprising: preprocessing radar echo image data and raindrop spectrometer data; quantitatively estimating radar reflectivity factors according to raindrop spectrometer data; extrapolating radar echo image data based on an optical flow method and a UNet network model, respectively, and extrapolating radar reflectivity factors based on an LSTM network model; dynamically regressing the extrapolated radar echo image data to obtain radar echo image data of a required forecast period; globally correcting the radar echo image data of a required forecast period according to the extrapolated radar reflectivity factors of the required forecast period; and forecasting rainfall according to the globally corrected radar echo image data. The present invention can make radar echo image data closer to near-ground rainfall information, improve the accuracy of rainfall forecasting, and improve adaptability to complex and changeable meteorological data.

Description

Technical Field

[0001] The invention relates to the technical field of meteorology and hydrology, and in particular to a method and system for short-term rainfall forecasting based on a raindrop spectrometer and a meteorological radar. Background Art

[0002] Traditional rainfall monitoring and forecasting mainly rely on measurements from automatic rain gauges. Although automatic rain gauges can provide ground-based measured data, their deployment density is difficult to meet the needs of high-precision monitoring, which limits their ability to provide continuous, real-time rainfall information over a large area. Although meteorological radars can cover a large area and provide high-altitude signal monitoring, their ability to extrapolate specific rainfall conditions close to the ground is limited, especially in complex terrain conditions, where their accuracy is significantly affected. Although raindrop spectrometers can provide detailed information on local rainfall characteristics, they are only applicable to specific locations due to their extremely limited monitoring range, making it difficult to make large-scale rainfall forecasts.

[0003] Therefore, there is an urgent need for a short-term rainfall forecasting method and system based on a raindrop spectrometer and a meteorological radar, which can make the radar echo image data closer to the near-ground rainfall information, improve the accuracy of rainfall forecasting and the adaptability to complex and changeable meteorological data. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a short-term rainfall forecasting method and system based on a raindrop spectrometer and a meteorological radar, which can make the radar echo image data closer to the near-ground rainfall information, improve the accuracy of rainfall forecasting and the adaptability to complex and changeable meteorological data.

[0005] The present invention provides a short-term rainfall forecasting method based on a raindrop spectrometer and a weather radar, comprising the following steps:

[0006] S1, obtaining radar echo image data and raindrop spectrometer data, and performing preprocessing;

[0007] S2, quantitatively estimating radar reflectivity factor based on preprocessed raindrop spectrometer data;

[0008] S3, constructing a first data set according to the preprocessed radar echo image data, constructing a second data set according to the radar reflectivity factor quantitatively estimated based on the preprocessed raindrop spectrometer data, using the first data set to train the UNet network model to obtain a trained UNet network model, and using the second data set to train the LSTM network model to obtain a trained LSTM network model;

[0009] S4, extrapolating the radar echo image data based on the optical flow method and the trained UNet network model respectively, and extrapolating the radar reflectivity factor quantitatively estimated based on the raindrop spectrometer data based on the trained LSTM network model;

[0010] S5. Perform dynamic regression on the extrapolated radar echo image data to obtain the radar echo image data of the required forecast period;

[0011] S6. performing global correction on the radar echo image data of the required forecast period according to the radar reflectivity factor of the required forecast period obtained by extrapolation;

[0012] S7. Predict rainfall based on the globally corrected radar echo image data.

[0013] Furthermore, in S1, radar echo image data and raindrop spectrometer data are obtained and preprocessed including:

[0014] Perform denoising, interpolation, coordinate transformation and data normalization processing on the acquired radar echo image data;

[0015] Calculate the raindrop number concentration based on the acquired raindrop spectrometer data and remove outliers;

[0016] The calculation formula of raindrop number concentration is as follows:

[0017] ;

[0018] Among them, N(D i ) represents the raindrop number concentration, D i represents the raindrop size corresponding to the discretization interval of the i-th raindrop size, j represents the j-th discretization interval of the final falling speed of the raindrop, i represents the i-th discretization interval of the raindrop size, n ij represents the number of raindrops corresponding to the discretization interval of the i-th raindrop particle size and the discretization interval of the j-th raindrop falling final velocity, A represents the sampling area, Δt represents the sampling time interval, V j represents the velocity corresponding to the discretized interval of the final velocity of the jth raindrop falling, ΔD i Represents the raindrop diameter gradient.

[0019] Furthermore, in S2, the calculation formula for quantitatively estimating the radar reflectivity factor based on the preprocessed raindrop spectrometer data is as follows:

[0020] ;

[0021] Where Z represents the radar reflectivity factor.

[0022] Furthermore, in S5, the extrapolated radar echo image data is dynamically regressed to obtain radar echo image data of the required forecast period, including:

[0023] S51, respectively comparing the radar echo image data extrapolated by the optical flow method and the trained UNet network model in the previous time period with the radar echo image data actually collected by the meteorological radar in the previous time period, and calibrating the regression weights of the UNet network model and the optical flow method based on the MSE loss function;

[0024] S52, according to the radar echo image data actually collected by the meteorological radar in the previous time period, the radar echo image data of the next time period is extrapolated based on the optical flow method and the trained UNet network model;

[0025] S53, according to the regression weights of the UNet network model and the optical flow method, weighted summing is performed on the radar echo image data of the next time period extrapolated by the UNet network model and the optical flow method to obtain the final radar echo image data of the next time period;

[0026] S54, repeat the above steps, and adjust the regression weights of the UNet network model and the optical flow method time step by time until the radar echo image data of the required forecast period is obtained.

[0027] Furthermore, in S51, the calculation formula of the MSE loss function is as follows:

[0028] ;

[0029] in, represents the dynamic regression result of radar echo image data extrapolated by UNet network model and optical flow method, y represents the radar echo image data actually collected by the meteorological radar, n represents the number of radar reflectivity factors, and k represents the kth radar reflectivity factor;

[0030] The expression of the dynamic regression result of radar echo image data extrapolated by UNet network model and optical flow method is:

[0031] ;

[0032] Among them, x UNet is the radar echo image extrapolated by the UNet network model, x OF is the radar echo image extrapolated by the optical flow method, c and d are the dynamically calibrated regression parameters, both in the range of [0,1]. The regression parameters are traversed and calibrated based on the range of values.

[0033] Further, in S6, performing global correction on the radar echo image data of the required forecast period according to the extrapolated radar reflectivity factor of the required forecast period includes:

[0034] S61, determining the spatial position of the raindrop spectrometer sampling point, and comparing it with the spatial position of the radar echo image data;

[0035] S62, comparing the radar reflectivity factor of the required forecast period obtained by extrapolating the LSTM network model with the radar reflectivity factor of the spatial position corresponding to the radar echo image data of the required forecast period to obtain a proportional relationship between the two;

[0036] S63. Globally scale the radar echo image data of the required forecast period according to the proportional relationship to obtain globally corrected radar echo image data.

[0037] Further, in S7, rainfall forecast is performed based on the globally corrected radar echo image data, and the calculation process is as follows:

[0038] ;

[0039] Among them, Z represents the radar reflectivity factor, R represents the inverted rainfall intensity, and a and b are empirical coefficients;

[0040] The values ​​of a and b are determined by the CTF error discriminant function:

[0041] ;

[0042] Among them, G q represents the rainfall intensity data measured by the automatic rain gauge station in the time period corresponding to the inverted rainfall intensity, q represents the station number of the automatic rain gauge station, and m represents the number of automatic rain gauge stations;

[0043] According to a set of a and b parameters with the smallest CTF error discriminant function, the rainfall intensity is inverted and the rainfall forecast is performed based on the globally corrected radar echo image data.

[0044] The present invention also provides a short-term rainfall forecasting system based on a raindrop spectrometer and a meteorological radar, which is used to implement any of the above-mentioned short-term rainfall forecasting methods based on a raindrop spectrometer and a meteorological radar, and includes the following modules:

[0045] A data acquisition module is used to acquire radar echo image data and raindrop spectrometer data and perform preprocessing; and quantitatively estimate the radar reflectivity factor based on the preprocessed raindrop spectrometer data;

[0046] A model training module, connected to the data acquisition module, is used to form a first data set according to the preprocessed radar echo image data, form a second data set according to the radar reflectivity factor quantitatively estimated based on the preprocessed raindrop spectrometer data, use the first data set to train the UNet network model to obtain a trained UNet network model, and use the second data set to train the LSTM network model to obtain a trained LSTM network model;

[0047] The prediction module is connected to the model training module and is used to predict the radar echo image data to be tested based on the optical flow method and the trained UNet network model, and to predict the radar reflectivity factor to be tested quantitatively estimated based on the raindrop spectrometer data to be tested based on the trained LSTM network model;

[0048] A calculation module, connected to the prediction module, for dynamically regressing the predicted radar echo image data;

[0049] A correction module, connected to the calculation module, is used to perform global correction on the radar echo image data after dynamic regression according to the predicted radar reflectivity factor;

[0050] The output module is connected with the correction module and is used for making rainfall forecast according to the radar echo image data after global correction.

[0051] The embodiments of the present invention have the following technical effects:

[0052] 1. The present invention couples the UNet network model and the optical flow method, which not only takes advantage of the deep learning model's prediction of nonlinear motion, but also takes advantage of the numerical calculation method's prediction clarity, thereby improving the accuracy of the extrapolated radar echo image. At the same time, the dynamic regression method is used to calibrate the weights of the UNet network model and the optical flow method in real time, which has strong adaptability to complex and changeable meteorological data.

[0053] 2. The present invention uses the extrapolated radar reflectivity factor to globally correct the extrapolated radar echo image, and uses the information collected by the raindrop spectrometer to correct the radar echo data, which can improve the accuracy of the echo image and make the corrected radar echo data closer to the near-ground rainfall information, thereby improving the accuracy of rainfall forecast.

[0054] 3. In the present invention, when inverting radar echo images according to the dynamic ZR relationship method, the information collected by the automatic rain gauge is included in the reference in real time by dynamically adjusting the a and b parameters, thereby enhancing the effective use of existing data, avoiding the influence of inversion errors caused by constant formulas and different regional characteristics, fully considering the influence of actual physical mechanisms on the ZR relationship, and improving the inversion accuracy of rainfall intensity information. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0056] Figure 1 It is a flow chart of a method for short-term rainfall forecasting based on raindrop spectrum and meteorological radar provided by an embodiment of the present invention;

[0057] Figure 2 It is a structural schematic diagram of a short-term rainfall forecasting system based on raindrop spectrum and meteorological radar provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the scope of protection of the present invention.

[0059] The present invention proposes a short-term rainfall forecasting method based on raindrop spectrum and meteorological radar. Figure 1 is a flow chart of a method for short-term rainfall forecasting based on raindrop spectrum and meteorological radar provided by an embodiment of the present invention, see Figure 1 , specifically including:

[0060] S1. Obtain radar echo image data and raindrop spectrometer data, and perform preprocessing.

[0061] Specifically, we first obtain multi-source data such as rainfall observation data and weather radar data based on the physical properties associated with the radar reflectivity factor and rainfall intensity, and then perform denoising, interpolation, coordinate transformation and data normalization on the acquired radar echo image data to ensure data quality and consistency.

[0062] Calculate the raindrop number concentration based on the acquired raindrop spectrometer data and remove outliers, wherein the acquired raindrop spectrometer data is discretized data;

[0063] The calculation formula of raindrop number concentration is as follows:

[0064] ;

[0065] Among them, N(D i ) represents the raindrop number concentration, D irepresents the raindrop size corresponding to the discretization interval of the i-th raindrop size, j represents the j-th discretization interval of the final falling speed of the raindrop, i represents the i-th discretization interval of the raindrop size, n ij represents the number of raindrops corresponding to the discretization interval of the i-th raindrop particle size and the discretization interval of the j-th raindrop falling final velocity, A represents the sampling area, Δt represents the sampling time interval, V j represents the velocity corresponding to the discretized interval of the final velocity of the jth raindrop falling, ΔD i Represents the raindrop diameter gradient.

[0066] S2. Quantitatively estimate the radar reflectivity factor based on the preprocessed raindrop spectrometer data.

[0067] Specifically, the calculation formula is as follows:

[0068] ;

[0069] Where Z represents the radar reflectivity factor.

[0070] S3. Construct a first data set according to the preprocessed radar echo image data, construct a second data set according to the radar reflectivity factor quantitatively estimated based on the preprocessed raindrop spectrometer data, use the first data set to train the UNet network model to obtain a trained UNet network model, and use the second data set to train the LSTM network model to obtain a trained LSTM network model.

[0071] Specifically, the first data set is constructed according to the preprocessed radar echo image data, and is divided into a data group and a label group according to the required forecast period. The data group is used as input and the label group is used as output to train the UNet network model. The second data set is processed in the same way to train the LSTM network model.

[0072] Exemplarily, the rainfall forecast model for the Liulin Basin is trained with a target forecast period of 30 minutes. 10,000 radar echo images of the Liulin Basin are collected with a collection frequency of 6 minutes, that is, 5 images for 30 minutes. The 10,000 images are interlaced and grouped into 5-frame images to obtain 9,995 data group images and label group images. The time of the data group images is 30 minutes earlier than that of the label group images, that is, 5 frames of images. For example, if the time of the first frame of the data group image is 10:00 on September 1, 2000, then the time of the first frame of the label group image is 10:30 on September 1, 2000, and so on.

[0073] The UNet network model is a U-shaped network structure model for semantic cutting, which includes three components: compression path, expansion path and jump connection. The image information first enters the compression path, where the model encodes and downsamples the image to form a feature map. In the expansion path, the model decodes and deconvolves the feature map, and cascades it with the feature map in the compression path through jump connections to repair some feature information lost in the convolution stage. The model is modified and applied to the extrapolation of radar echo images through multiple frames of continuous radar echo images, so as to make rainfall forecasts. The LSTM network model, namely the Long Short-Term Memory network, is a special type of recurrent neural network (RNN). It is designed to deal with the "gradient disappearance" or "gradient explosion" problems encountered by traditional RNNs when processing sequence data. The radar reflectivity factor quantitatively estimated by the raindrop spectrum over a period of time is sorted and preprocessed to convert it into a form suitable for model input, that is, a series of continuous time point data. Then, the LSTM network model is trained with these data to learn to capture the dynamic changes of the radar reflectivity factor pattern from the input sequence. After training, the model can predict the radar reflectivity factor of the raindrop spectrum quantitatively estimated in the next period of time (such as the next hour) based on the radar reflectivity factor of the recent raindrop spectrum quantitative estimation (such as the data of the past 3 hours). LSTM effectively retains and utilizes the long-term dependencies in historical data through its internal memory units and gating mechanisms, thereby achieving accurate prediction of future rainfall.

[0074] S4. The radar echo image data is extrapolated based on the optical flow method and the trained UNet network model respectively, and the radar reflectivity factor quantitatively estimated based on the raindrop spectrometer data is extrapolated based on the trained LSTM network model.

[0075] Specifically, the present invention couples the UNet network model and the optical flow method, which not only utilizes the prediction advantage of the deep learning model for nonlinear motion, but also utilizes the advantage of the numerical calculation method in forecasting clarity, thereby improving the accuracy of the extrapolated radar echo image. At the same time, the dynamic regression method is used to calibrate the weights of the UNet network model and the optical flow method in real time, which has strong adaptability to complex and changeable meteorological data.

[0076] The present invention also uses the LSTM network model to quantitatively estimate the radar reflectivity factor of the raindrop spectrum. The advantage of the LSTM model in extrapolating and predicting the radar reflectivity factor is that it can effectively capture the dependencies within a long time span, filter out irrelevant information and retain key long-term patterns through unique memory units and gating mechanisms. This feature enables LSTM to more accurately understand and predict the changing characteristics of the radar reflectivity factor, especially when processing nonlinear, complex and intermittent meteorological data. Therefore, the LSTM model can provide more accurate extrapolation of the radar reflectivity factor than traditional methods.

[0077] S5. Perform dynamic regression on the extrapolated radar echo image data to obtain the radar echo image data of the required forecast period.

[0078] The specific steps include:

[0079] S51. Compare the radar echo image data extrapolated by the optical flow method and the trained UNet network model in the previous time period with the radar echo image data actually collected by the meteorological radar in the previous time period, and calibrate the regression weights of the UNet network model and the optical flow method based on the MSE loss function.

[0080] Specifically, the calculation formula of the MSE loss function is as follows:

[0081] ;

[0082] in, represents the dynamic regression result of radar echo image data extrapolated by UNet network model and optical flow method, y represents the radar echo image data actually collected by the meteorological radar, n represents the number of radar reflectivity factors, and k represents the kth radar reflectivity factor;

[0083] The expression of the dynamic regression result of radar echo image data extrapolated by UNet network model and optical flow method is:

[0084] ;

[0085] Among them, x UNet is the radar echo image extrapolated by the UNet network model, x OF is the radar echo image extrapolated by the optical flow method, c and d are the dynamically calibrated regression parameters, both in the range of [0,1]. The regression parameters are traversed and calibrated based on the range of values.

[0086] S52. According to the radar echo image data actually collected by the meteorological radar in the previous time period, the radar echo image data for the next time period is extrapolated based on the optical flow method and the trained UNet network model.

[0087] S53. According to the regression weights of the UNet network model and the optical flow method, weighted summation is performed on the radar echo image data of the next time period extrapolated by the UNet network model and the optical flow method to obtain the final radar echo image data of the next time period.

[0088] S54, repeat the above steps, and adjust the regression weights of the UNet network model and the optical flow method time step by time until the radar echo image data of the required forecast period is obtained.

[0089] For example, due to the different calculation logics of the optical flow method and the UNet network model, the number of input images required by the two is generally inconsistent (the number of input images for different forecast periods is also different). Suppose we need to forecast the radar echo image of the Liulin Experimental Basin from 2:00 to 2:30 p.m. First, the real radar echo images of the Liulin Experimental Basin collected by the meteorological radar from 1:24 to 1:30 p.m. and 1:00 to 1:30 p.m. are input to the optical flow method and the UNet network model respectively. The two models extrapolate the radar echo images to obtain the radar echo images of the Liulin Experimental Basin from 1:30 to 2:00 p.m. The radar echo images of the Liulin Experimental Basin from 1:30 to 2:00 p.m. extrapolated by the UNet network model and the optical flow method are compared with the real radar echo images of the Liulin Experimental Basin from 1:30 to 2:00 p.m. collected by the meteorological radar, and the MSE loss function is used to calibrate the regression weights of the UNet network model and the optical flow method. Then, the real radar echo images of the Liulin Experimental Basin collected by the meteorological radar from 1:54 to 2:00 p.m. and 1:30 to 2:00 p.m. were input into the UNet network model and the optical flow method respectively. The two models were extrapolated to obtain the radar echo images of the Liulin Experimental Basin from 2:00 to 2:30 p.m. The extrapolated results of the two models were weighted and summed according to the previously calibrated regression weights to obtain the final radar echo image data from 2:00 to 2:30 p.m. The subsequent time series were deduced in this way to achieve the effect of dynamic regression.

[0090] Furthermore, each time period needs to calibrate the regression weights of the UNet model and the optical flow method based on the extrapolated radar echo image data of the UNet model and the optical flow method of the previous time period, that is, the regression weights change dynamically, with better spatiotemporal accuracy.

[0091] S6. Performing a global correction on the radar echo image data of the required forecast period according to the extrapolated radar reflectivity factor of the required forecast period.

[0092] The specific steps include:

[0093] S61. Determine the spatial position of the raindrop spectrometer sampling point and compare it with the spatial position of the radar echo image data.

[0094] S62. Compare the radar reflectivity factor of the required forecast period obtained by extrapolating the LSTM network model with the radar reflectivity factor of the spatial position corresponding to the radar echo image data of the required forecast period to obtain a proportional relationship between the two.

[0095] S63. Globally scale the radar echo image data of the required forecast period according to the proportional relationship to obtain globally corrected radar echo image data.

[0096] For example, the radar reflectivity factor extrapolated by the LSTM network model is used to perform global correction on the radar echo image of the Liulin Experimental Basin from 2:00 to 2:30 pm obtained by dynamic regression. First, the radar reflectivity factor of the raindrop spectrum quantitatively estimated from the Liulin Experimental Basin from 1:30 to 2:00 pm is input into the LSTM network model, and the radar reflectivity factor of the Liulin Experimental Basin from 2:00 to 2:30 pm is extrapolated. The radar reflectivity factor of the Liulin Experimental Basin from 2:00 to 2:30 pm extrapolated by the LSTM network model is compared with the radar reflectivity factor of the corresponding spatial point of the radar echo image of the Liulin Experimental Basin from 2:00 to 2:30 pm obtained by dynamic regression, and the dynamically regressed radar echo image is globally scaled to achieve global correction.

[0097] The present invention utilizes the radar reflectivity factor obtained by extrapolation to globally correct the radar echo image obtained by extrapolation, and utilizes the information collected by the raindrop spectrometer to correct the radar echo data, thereby improving the accuracy of the echo image and making the corrected radar echo data closer to the near-ground rainfall information, thereby improving the accuracy of rainfall forecasting.

[0098] S7. Predict rainfall based on the globally corrected radar echo image data.

[0099] Specifically, the calculation process is as follows:

[0100] ;

[0101] Wherein, Z represents the radar reflectivity factor, R represents the inverted rainfall intensity, and a and b are empirical coefficients; illustratively, the value range of a can be set to [1,300,1], where [1,300] is the value range and 1 is the step size, and the value range of b can be set to [0.5,5,0.1], where [0.5,5] is the value range and 0.1 is the step size;

[0102] In each stage of forecasting and inverting rainfall intensity based on actual radar echo images, the radar echo images of the previous period are first calculated and inverted through the range of a and b values, and compared with the measured data of the automatic rain gauge station, and the closest set of a and b values ​​are selected as the values ​​for calculation and inversion in this period to achieve the effect of dynamic value selection. The CTF (Cost Function) error discriminant function is used to compare the calculation results during the traversal period with the measured data of the automatic rain gauge station to achieve the optimal fitting effect; the values ​​of a and b are determined by the CTF error discriminant function, and the calculation formula is as follows:

[0103] ;

[0104] Among them, G q represents the rainfall intensity data measured by the automatic rain gauge station in the time period corresponding to the inverted rainfall intensity, q represents the station number of the automatic rain gauge station, and m represents the number of automatic rain gauge stations;

[0105] According to a set of a and b parameters with the smallest CTF error discriminant function, the rainfall intensity is inverted and the rainfall forecast is performed based on the globally corrected radar echo image data.

[0106] For example, if there are 5 rain gauges in the Liulin Basin, then m=5. The rainfall intensity information at 10:30 on September 1, 2020 is forecasted, with a forecast period of 30 minutes. First, the accuracy of all a and b values ​​of the echo image at 10:00 on September 1, 2020 is traversed, that is, the rainfall intensity information of 300×45 groups of 5 rain gauges is obtained, and the CTF error function is calculated for all of them. The group of a and b parameters with the smallest CTF value is selected as the ZR relationship for inverting the rainfall intensity information at 10:30 on September 1 for inversion calculation.

[0107] In the present invention, when inverting radar echo images according to the dynamic ZR relationship method, the information collected by the automatic rain gauge is taken into reference in real time by dynamically adjusting the a and b parameters, thereby enhancing the effective use of existing data, avoiding the influence of inversion errors caused by constant formulas and different regional characteristics, fully considering the influence of actual physical mechanisms on the ZR relationship, and improving the inversion accuracy of rainfall intensity information.

[0108] The present invention also proposes a short-term rainfall forecasting system based on raindrop spectrum and meteorological radar. Figure 2 is a structural schematic diagram of a short-term rainfall forecasting system based on raindrop spectrum and meteorological radar provided by an embodiment of the present invention, which is used to implement any of the above-mentioned short-term rainfall forecasting methods based on raindrop spectrum and meteorological radar, see Figure 2 , including the following modules:

[0109] A data acquisition module is used to acquire radar echo image data and raindrop spectrometer data and perform preprocessing; and quantitatively estimate the radar reflectivity factor based on the preprocessed raindrop spectrometer data;

[0110] A model training module, connected to the data acquisition module, is used to form a first data set according to the preprocessed radar echo image data, form a second data set according to the radar reflectivity factor quantitatively estimated based on the preprocessed raindrop spectrometer data, use the first data set to train the UNet network model to obtain a trained UNet network model, and use the second data set to train the LSTM network model to obtain a trained LSTM network model;

[0111] The prediction module is connected to the model training module and is used to predict the radar echo image data to be tested based on the optical flow method and the trained UNet network model, and to predict the radar reflectivity factor to be tested quantitatively estimated based on the raindrop spectrometer data to be tested based on the trained LSTM network model;

[0112] A calculation module, connected to the prediction module, for dynamically regressing the predicted radar echo image data;

[0113] A correction module, connected to the calculation module, is used to perform global correction on the radar echo image data after dynamic regression according to the predicted radar reflectivity factor;

[0114] The output module is connected with the correction module and is used for making rainfall forecast according to the radar echo image data after global correction.

[0115] It should be noted that the terms used in the present invention are only for describing specific embodiments, rather than limiting the scope of the present application. As shown in the present specification, unless the context clearly indicates an exception, the words "one", "a", "a kind of" and / or "the" do not specifically refer to the singular, but may also include the plural. The terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method or device. In the absence of more restrictions, the elements defined by the sentence "include one..." do not exclude the presence of other identical elements in the process, method or device including the elements.

[0116] It should also be noted that the terms "center", "up", "down", "left", "right", "vertical", "horizontal", "inside", "outside", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. Unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", etc. should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be a connection between the two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the technical solutions of the embodiments of the present invention.

Claims

1. A short-term rainfall forecasting method based on a raindrop spectrometer and a weather radar, characterized in that: The steps include: S1, obtaining radar echo image data and raindrop spectrometer data, and performing preprocessing; S2, quantitatively estimating radar reflectivity factor based on preprocessed raindrop spectrometer data; S3, constructing a first data set according to the preprocessed radar echo image data, constructing a second data set according to the radar reflectivity factor quantitatively estimated based on the preprocessed raindrop spectrometer data, using the first data set to train the UNet network model to obtain a trained UNet network model, and using the second data set to train the LSTM network model to obtain a trained LSTM network model; S4, extrapolating the radar echo image data based on the optical flow method and the trained UNet network model respectively, and extrapolating the radar reflectivity factor quantitatively estimated based on the raindrop spectrometer data based on the trained LSTM network model; S5. Perform dynamic regression on the extrapolated radar echo image data to obtain the radar echo image data of the required forecast period; S6. performing global correction on the radar echo image data of the required forecast period according to the radar reflectivity factor of the required forecast period obtained by extrapolation; S7. Predict rainfall based on the globally corrected radar echo image data.

2. The method for short-term rainfall forecasting based on raindrop spectrometer and weather radar according to claim 1, characterized in that: In S1, radar echo image data and raindrop spectrometer data are obtained and preprocessed, including: Perform denoising, interpolation, coordinate transformation and data normalization processing on the acquired radar echo image data; Calculate the raindrop number concentration based on the acquired raindrop spectrometer data and remove outliers; The calculation formula of raindrop number concentration is as follows: ； Among them, N(D i ) represents the raindrop number concentration, D i represents the raindrop size corresponding to the discretization interval of the i-th raindrop size, j represents the j-th discretization interval of the final falling speed of the raindrop, i represents the i-th discretization interval of the raindrop size, n ij represents the number of raindrops corresponding to the discretization interval of the i-th raindrop particle size and the discretization interval of the j-th raindrop falling final velocity, A represents the sampling area, Δt represents the sampling time interval, V j represents the velocity corresponding to the discretized interval of the final velocity of the jth raindrop falling, ΔD i Represents the raindrop diameter gradient.

3. The method for short-term rainfall forecasting based on raindrop spectrometer and weather radar according to claim 2 is characterized in that: In S2, the calculation formula for quantitatively estimating the radar reflectivity factor based on the pre-processed raindrop spectrometer data is as follows: ； Where Z represents the radar reflectivity factor.

4. The method for short-term rainfall forecasting based on raindrop spectrometer and weather radar according to claim 1, characterized in that: In S5, the externally extrapolated radar echo image data is dynamically regressed to obtain radar echo image data of the required forecast period, including: S51, respectively comparing the radar echo image data extrapolated by the optical flow method and the trained UNet network model in the previous time period with the radar echo image data actually collected by the meteorological radar in the previous time period, and calibrating the regression weights of the UNet network model and the optical flow method based on the MSE loss function; S52, according to the radar echo image data actually collected by the meteorological radar in the previous time period, the radar echo image data of the next time period is extrapolated based on the optical flow method and the trained UNet network model; S53, according to the regression weights of the UNet network model and the optical flow method, weighted summing is performed on the radar echo image data of the next time period extrapolated by the UNet network model and the optical flow method to obtain the final radar echo image data of the next time period; S54, repeat the above steps, and adjust the regression weights of the UNet network model and the optical flow method time step by time until the radar echo image data of the required forecast period is obtained.

5. The method for short-term rainfall forecasting based on raindrop spectrometer and weather radar according to claim 4 is characterized in that: In S51, the calculation formula of the MSE loss function is as follows: ； in, represents the dynamic regression result of radar echo image data extrapolated by UNet network model and optical flow method, y represents the radar echo image data actually collected by the meteorological radar, n represents the number of radar reflectivity factors, and k represents the kth radar reflectivity factor; The expression of the dynamic regression result of radar echo image data extrapolated by UNet network model and optical flow method is: ； Among them, x UNet is the radar echo image extrapolated by the UNet network model, x OF is the radar echo image extrapolated by the optical flow method, c and d are the dynamically calibrated regression parameters, both in the range of [0,1]. The regression parameters are traversed and calibrated based on the range of values.

6. The method for short-term rainfall forecasting based on raindrop spectrometer and weather radar according to claim 5, characterized in that: In S6, performing global correction on the radar echo image data of the required forecast period according to the extrapolated radar reflectivity factor of the required forecast period includes: S61, determining the spatial position of the raindrop spectrometer sampling point, and comparing it with the spatial position of the radar echo image data; S62, comparing the radar reflectivity factor of the required forecast period obtained by extrapolating the LSTM network model with the radar reflectivity factor of the spatial position corresponding to the radar echo image data of the required forecast period to obtain a proportional relationship between the two; S63. Globally scale the radar echo image data of the required forecast period according to the proportional relationship to obtain globally corrected radar echo image data.

7. The method for short-term rainfall forecasting based on raindrop spectrometer and weather radar according to claim 6, characterized in that: In S7, rainfall forecast is performed based on the radar echo image data after global correction, and the calculation process is as follows: ； Among them, Z represents the radar reflectivity factor, R represents the inverted rainfall intensity, and a and b are empirical coefficients; The values ​​of a and b are determined by the CTF error discriminant function: ； Among them, G q represents the rainfall intensity data measured by the automatic rain gauge station in the time period corresponding to the inverted rainfall intensity, q represents the station number of the automatic rain gauge station, and m represents the number of automatic rain gauge stations; According to a set of a and b parameters with the smallest CTF error discriminant function, the rainfall intensity is inverted and the rainfall forecast is performed based on the globally corrected radar echo image data.

8. A short-term rainfall forecasting system based on a raindrop spectrometer and a weather radar, used to implement a short-term rainfall forecasting method based on a raindrop spectrometer and a weather radar as described in any one of claims 1 to 7, characterized in that: Includes the following modules: A data acquisition module is used to acquire radar echo image data and raindrop spectrometer data and perform preprocessing; and quantitatively estimate the radar reflectivity factor based on the preprocessed raindrop spectrometer data; A model training module, connected to the data acquisition module, for forming a first data set according to the preprocessed radar echo image data, forming a second data set according to the radar reflectivity factor quantitatively estimated based on the preprocessed raindrop spectrometer data, using the first data set to train the UNet network model to obtain a trained UNet network model, and using the second data set to train the LSTM network model to obtain a trained LSTM network model; A prediction module, connected to the model training module, is used to predict the radar echo image data to be tested based on the optical flow method and the trained UNet network model, and to predict the radar reflectivity factor to be tested quantitatively estimated based on the raindrop spectrometer data to be tested based on the trained LSTM network model; A calculation module, connected to the prediction module, for dynamically regressing the predicted radar echo image data; A correction module, connected to the calculation module, for globally correcting the radar echo image data after dynamic regression according to the predicted radar reflectivity factor; The output module is connected to the correction module and is used to make rainfall forecast according to the radar echo image data after global correction.

Images