A short-term forecast method for severe convection based on deep learning and vertical profile of polarization quantity

Abstract

The present invention provides a short-term forecast method for severe convection based on deep learning and vertical profile of polarization. According to the three-dimensional grid data of polarization and wind field of contour surface at preset time intervals within the radar network range in the historical time period, convective clouds are identified; the convective clouds are tracked by the CLTREC method, and the front-back matching relationship of the convective clouds is established, and the time-series grid point position of the convective clouds in the historical time period is determined according to the front-back matching relationship; according to the time-series grid point position of the convective cloud, the time-series characteristics of the vertical profile of the polarization of the convective cloud grid point are sampled from the three-dimensional grid point data, and the actual posterior probability of the time-series characteristics of the vertical profile of the polarization of the convective cloud grid point is determined; the actual posterior probability is compared with the probability threshold, and whether severe convection will occur in the future time period is determined according to the comparison result, and the probability threshold is optimized by the deep learning model. The present invention improves the forecast lead time and forecast accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological forecasting, and in particular to a method for short-term forecasting of severe convection based on deep learning and vertical profiles of polarization quantities. Background Art

[0002] Severe convective weather refers to severe convection during the flood season that produces disastrous weather such as thunderstorms and gales (>=17.2 m / s) or short-term heavy rainfall (>=20 mm / h). Severe convection is a difficult research topic in the atmospheric science community. Research on the predictability of severe convection shows that the current global operational numerical forecasting model and mesoscale numerical model have very limited forecasting capabilities for extreme severe convection. Due to the obvious sudden, localized, and disaster-causing characteristics of extreme severe convection in meteorological operations, the actual effective warning time is still generally less than 1 hour. Summary of the invention

[0003] The present invention provides a severe convection short-term forecasting method based on deep learning and vertical profile of polarization quantity, so as to solve the defects of poor forecast timeliness and accuracy in the prior art and improve the forecast lead time and forecast accuracy.

[0004] The present invention provides a severe convection short-term prediction method based on deep learning and polarization vertical profile, comprising:

[0005] Convective clouds are identified based on the three-dimensional grid data of the polarization quantity and wind field of the contour surface at preset time intervals within the radar network range in the historical time period;

[0006] Tracking the convective clouds using the CLTREC method, establishing a front-to-back matching relationship of the convective clouds, and determining the temporal grid point positions of the convective clouds in the historical time period according to the front-to-back matching relationship;

[0007] According to the time-series grid point positions of the convective cloud, sampling the time-series features of the vertical profile of the polarization amount of the convective cloud grid points from the three-dimensional grid point data, and determining the actual posterior probability of the time-series features of the vertical profile of the polarization amount of the convective cloud grid points;

[0008] The actual posterior probability is compared with a probability threshold, and whether severe convection will occur in a future time period is determined based on the comparison result. The probability threshold is optimized using a deep learning model.

[0009] According to a method for short-term severe convection forecasting based on deep learning and vertical profile of polarization quantity provided by the present invention, the convective cloud is tracked by the CLTREC method, and a front-back matching relationship of the convective cloud is established, and the method also includes:

[0010] Perform vertical motion salient feature position correction for each tracking point position, and correct the position P VM The formula is as follows:

[0011]

[0012]

[0013]

[0014] in, It means that the initial position (X, Y) is used as the center, and the grid-by-grid search based on the differential propagation phase shift rate K is performed within the preset radius area. DP Calculated vertical liquid water content The position of the maximum value, the initial position is the grid point corresponding to the current convective cloud, cappiNum is the highest layer of the grid point, is the moving characteristics of convective clouds at adjacent moments The rate of change.

[0015] According to a method for short-term severe convection prediction based on deep learning and vertical profile of polarization quantity provided by the present invention, the actual posterior probability of the time series characteristics of the vertical profile of polarization quantity of the convective cloud grid is determined, including:

[0016] Establish a dataset of time series characteristics of vertical profiles of polarization quantity that characterize the precursors of severe convection;

[0017] Using the maximum likelihood estimation method to perform piecewise Gaussian function fitting on the time series characteristics of the vertical profile of the polarization quantity in the data set, a priori probability of the occurrence of strong convection within a preset time before the occurrence of the strong convection is constructed;

[0018] The prior probability is converted by using a priori to posterior probability conversion method to obtain the actual posterior probability of the convective cloud grid point.

[0019] According to a severe convection short-term prediction method based on deep learning and polarization vertical profile provided by the present invention, the time series characteristics of each polarization vertical profile and the actual posterior probability of each polarization vertical profile time series characteristics are a two-dimensional matrix;

[0020] The horizontal direction of the two-dimensional matrix is ​​the time dimension, and the vertical direction is the height dimension.

[0021] According to a method for short-term severe convection prediction based on deep learning and vertical profile of polarization quantity provided by the present invention, the actual posterior probability is compared with the probability threshold, and whether severe convection will occur in the future time period is determined according to the comparison result, including:

[0022] Calculate the percentage of actual posterior probabilities greater than the probability threshold in the two-dimensional matrix of actual posterior probabilities of the time series characteristics of the vertical profiles of each polarization quantity;

[0023] When the percentage is greater than a preset threshold, determining that the time series characteristics of the vertical profiles of the polarization quantities have precursor characteristics of severe convection;

[0024] When the time series characteristics of the vertical profiles of all polarization quantities have the precursory characteristics of severe convection, it is determined that severe convection will occur in the future time period.

[0025] According to a severe convection short-term forecasting method based on deep learning and vertical profile of polarization quantity provided by the present invention, the step of optimizing the probability threshold using the deep learning model includes:

[0026] Determine the actual posterior probability of the time series feature of the vertical profile of the polarization quantity in the time series feature training set of the vertical profile of the polarization quantity;

[0027] Performing Gaussian fitting on the actual posterior probability of the time series characteristics of the vertical profile of the polarization quantity in the training set to obtain a Gaussian curve;

[0028] Determine the probability difference between the left end of the Gaussian curve and the variance and twice the variance, and determine the search range of the probability threshold according to the two probability differences;

[0029] The actual posterior probability of the time series feature of the vertical profile of the polarization quantity in the training set is calculated through multiple convolution layers along the time dimension to obtain a posterior probability convolution index;

[0030] Utilizing the posterior probability convolution index of each convolution layer in the training set, calculating the optimal risk score corresponding to each convolution layer under the specified probability threshold within the search range;

[0031] The convolution layer corresponding to the maximum value in the optimal risk score is used as the optimal convolution layer, and the probability threshold corresponding to the maximum value in the optimal risk score is used as the optimal probability threshold.

[0032] According to a severe convection short-term forecasting method based on deep learning and vertical profile of polarization quantity provided by the present invention, the width of the convolution window of the multi-layer convolution layer is different, and the length is equal to the maximum height of the convective cloud grid.

[0033] According to a severe convection short-term forecasting method based on deep learning and polarization quantity vertical profile provided by the present invention, the time series characteristics of the polarization quantity vertical profile in the data set include physical and dynamic vertical profile time series characteristics, and the physical and dynamic vertical profile time series characteristics include one or more of a meso-γ scale horizontal reflectivity factor time series vertical profile, a meso-γ scale differential reflectivity time series vertical profile, a meso-γ scale differential propagation phase shift rate time series vertical profile, a meso-β scale vertical wind shear time series vertical profile, a meso-β scale jet stream characteristic time series vertical profile, a meso-β scale vorticity characteristic time series vertical profile, and a meso-β scale divergence characteristic time series vertical profile.

[0034] According to a severe convection short-term prediction method based on deep learning and vertical profile of polarization quantity provided by the present invention, the polarization quantity of the contour surface includes one or more of the horizontal reflectivity factor, the differential reflectivity and the differential propagation phase shift rate;

[0035] The three-dimensional grid data of the contour surface wind field includes one or more of the three-dimensional grid data of the contour surface vertical wind shear, the three-dimensional grid data of the contour surface vorticity, the three-dimensional grid data of the contour surface divergence and the three-dimensional grid data of the contour surface wind speed jet.

[0036] According to a method for short-term severe convection forecasting based on deep learning and vertical profiles of polarization provided by the present invention, before identifying convective clouds according to the three-dimensional grid data of polarization of contour surfaces and wind fields at preset time intervals within a historical time period within the radar networking range, the method further includes:

[0037] The S-band dual-polarization radar data in the historical time period within the radar network is quality controlled to generate polarization data of a single radar three-dimensional grid point;

[0038] Generate vertical wind profile data representing atmospheric environment wind field information in a historical time period within the radar network range by using a speed and azimuth display method;

[0039] The polarization data and vertical wind profile data of all radars within the radar networking range are gridded using a distance exponential weighted jigsaw method to obtain three-dimensional grid data of the polarization and wind field on the contour surface.

[0040] The method for short-term severe convection forecasting based on deep learning and vertical profile of polarization quantity provided by the present invention improves the ability to identify severe convection in advance by integrating the time series feature information of the vertical profile of polarization quantity characterizing the precursor of severe convection into the short-term severe convection forecasting; and optimizes the probability threshold by combining the deep learning method to improve the forecast accuracy of severe convection, thereby achieving a comprehensive improvement in the short-term severe convection forecasting capability. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0042] Figure 1 It is a flow chart of a method for short-term severe convection forecasting based on deep learning and vertical profile of polarization quantity provided by the present invention;

[0043] Figure 2It is a schematic diagram of short-term heavy rainfall forecast and actual situation comparison of the severe convection short-term forecast method based on deep learning and vertical profile of polarization quantity provided by the present invention. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are 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 are within the scope of protection of the present invention.

[0045] The concern and difficulty of short-term nowcasting is how to capture favorable precursor information in time before the occurrence of severe convection, especially when the convective weather system has just been generated or developed, or captures early signals at an earlier stage, which has very important scientific significance and application value. Many studies have shown that indicative cloud macro microphysical and dynamic characteristic signals are integrated into forecasting or early warning models to enhance the early forecast and early warning capabilities of extremely severe convection. In addition, there are significant differences in the microphysical characteristics of precipitation in different types of convective systems. Even for specific weather systems, the microphysical structure characteristics of convection at different development stages and in different regions will also change significantly. Therefore, the present invention integrates the precursor characteristic information of the vertical profile of the polarization quantity of severe convection in the severe convection short-term forecast model, thereby providing the possibility for constructing a quantitative early warning model.

[0046] The generation of severe convection requires strong vertical movement, so its vertical scale characteristics should be more significant than the horizontal scale. However, most of the deep learning short-term forecasting methods only rely on the horizontal echo distribution information. Therefore, although the generation and evolution effect of severe convection (>40dBZ) has been achieved to a certain extent, the improvement of TS (Threat Score) is still limited. In addition, the horizontal convolution processing of the Convolutional Neural Network (CNN) represented by UNet can effectively capture the multi-scale evolution characteristics of severe convection, but inappropriate convolution scales may also affect the integrity of the characteristics, thereby affecting the model performance. Therefore, in order to improve the short-term forecasting capability of severe convection, the present invention refers to the mechanism of severe convection for the design of the deep learning model, focuses on using the physical and dynamic vertical structure time series characteristics that characterize the precursors of severe convection as model parameters, and then uses the deep learning method to adaptively learn the optimal short-term business threshold from the multimodal model with more significant characteristics.

[0047] The present invention focuses on the research on the precursor characteristic signals of multi-source vertical profile characteristics before the occurrence and development of severe convection, develops a short-term severe convection forecasting method based on the vertical profile characteristics of the polarization quantity, and combines the deep learning method to optimize the threshold, so as to further improve the forecast lead time and forecast accuracy.

[0048] The difficulty of the present invention is:

[0049] 1. How to integrate the time series characteristic information of the vertical profile of polarization quantity that represents the precursor of severe convection into the short-term forecast model, thereby improving the ability to identify severe convection in advance.

[0050] 2. How to construct a deep learning method to optimize the short-term forecast model that integrates the time series characteristics of the vertical profile of the polarization quantity, and thus provide a digital basis for deep optimization thresholds.

[0051] 3. How to ensure that deep learning training is stable and converges to the optimal results through reasonable constraints on the physical and dynamic mechanism conditions of severe convection precursors.

[0052] Combine the following Figure 1 The present invention describes a severe convection short-term prediction method based on deep learning and polarization vertical profile, comprising:

[0053] Step 101, identifying convective clouds according to the three-dimensional grid data of the polarization amount of the contour surface and the wind field at preset time intervals within the radar network range in the historical time period;

[0054] Step 102, using the CLTREC method to track the convective cloud, establish a front-to-back matching relationship of the convective cloud, and determine the time-series grid point position of the convective cloud in the historical time period according to the front-to-back matching relationship;

[0055] Step 103, sampling the vertical profile time series characteristics of the polarization quantity of the convective cloud grid points from the three-dimensional grid point data according to the time series grid point positions of the convective cloud, and determining the actual posterior probability of the vertical profile time series characteristics of the polarization quantity of the convective cloud grid points;

[0056] Step 104, comparing the actual posterior probability with a probability threshold, and determining whether severe convection will occur in a future time period based on the comparison result, wherein the probability threshold is optimized using a deep learning model.

[0057] For example, based on the polarization quantity of the contour surface and the three-dimensional grid data of the wind field every 6 minutes in the previous hour within the network range, the convection identification method using the echo intensity and gradient characteristics on the radar contour surface can be used to identify the convective cloud area.

[0058] The time series probability of severe convective properties in the future hourly is determined for each grid point in the identified convective cloud, thereby determining whether severe convective disasters will occur in the next hour. Finally, the location of severe convective disasters in the future is determined by the semi-Lagrangian extrapolation method.

[0059] The specific steps of determining the time series probability of strong convective attributes include: taking the three-dimensional grid data of every 6 minutes within the network range as the benchmark, firstly taking the current convective cloud grid point to be identified as the starting point and the initial position; then using the CLTREC method to track the convective cloud cluster, establishing the front and back matching relationship of the convective cloud cluster, and then searching for the convective cloud time series grid point position in the past hour; finally, performing vertical profile sampling of 5X5 grid points for the 6-minute interval convective cloud grid points tracked within 1 hour, and then generating the time series feature data of the vertical profile of the polarization of the current convective cloud grid point in the past hour (two-dimensional matrix, the x direction is the time dimension; the y direction is the height dimension); then according to the baseline posterior probability PRO of the time series feature variable of the vertical profile of the polarization of the corresponding preset time length S , convert to obtain the actual posterior probability PRO of the vertical profile time series characteristics of the polarization quantity of the current convective cloud grid point in the past hour O (two-dimensional matrix, x direction is time dimension; y direction is height dimension); the actual posterior probability PRO O The result is compared with the probability threshold, and it is determined whether a severe convective disaster will occur in the future based on the comparison result.

[0060] The CLTREC method is based on the traditional COTREC method, adding the echo intensity continuity constraint check and vector total variation correction at adjacent moments, so as to make the inverted radar echo motion vector field more continuous. Therefore, this embodiment adopts this method to generate regional echo tracking grid point motion vector field data.

[0061] Due to the baseline posterior probability PRO of the time series characteristic variable of the vertical profile of the polarization quantity S It is not guaranteed to always exist in Gaussian distribution. The actual distribution may have multi-peak characteristics. Therefore, the probability threshold PRO TH This embodiment uses a deep learning model to adjust the probability threshold PRO TH Perform tuning.

[0062] This embodiment improves the ability to identify severe convection in advance by integrating the time series characteristic information of the vertical profile of the polarization quantity that characterizes the precursor of severe convection into the short-term severe convection forecast; and combines the deep learning method to optimize the probability threshold to improve the forecast accuracy of severe convection, thereby achieving a comprehensive improvement in the short-term severe convection forecast capability.

[0063] On the basis of the above embodiment, the present embodiment uses the CLTREC method to track the convective cloud and establishes the front-back matching relationship of the convective cloud, and also includes:

[0064] Perform vertical motion salient feature position correction for each tracking point position, and correct the position P VM The formula is as follows:

[0065]

[0066]

[0067]

[0068] in, It means that the initial position (X, Y) is used as the center, and a grid-by-grid search is performed within a preset radius area (such as a 40km radius area) based on the differential propagation phase shift rate K. DP Calculated vertical liquid water content (i.e., accumulated from the first layer of the contour surface to the highest layer cappiNum) where the maximum value is located. The initial position is the grid point corresponding to the current convective cloud, and cappiNum is the highest layer of the grid point. is the moving characteristics of convective clouds at adjacent moments (calculated based on the CLTREC method) The rate of change.

[0069] On the basis of the above embodiment, in this embodiment, determining the actual posterior probability of the vertical profile time series characteristics of the polarization amount of the convective cloud grid point includes:

[0070] Establish a dataset of time series characteristics of vertical profiles of polarization quantity that characterize the precursors of severe convection;

[0071] Using the maximum likelihood estimation method to perform piecewise Gaussian function fitting on the time series characteristics of the vertical profile of the polarization quantity in the data set, a priori probability of the occurrence of strong convection within a preset time before the occurrence of the strong convection is constructed;

[0072] The prior probability is converted by using a priori to posterior probability conversion method to obtain the actual posterior probability of the convective cloud grid point.

[0073] Each valid time series feature sample in the polarization vertical profile time series feature data set is processed as follows:

[0074] Taking the 6-minute three-dimensional gridded data within the network range as the benchmark, the position of the automatic station with a live label is first taken as the initial position; then the convective cloud cluster is tracked using the CLTREC method to establish the front and back matching relationship of the convective cloud cluster, and at the same time, the vertical motion significant feature position correction is performed on each tracking point position, and then the convective cloud time series grid position within 3 hours before the occurrence of a severe convective disaster that represents the characteristics of convective propagation is searched out; finally, the vertical profile of the convective cloud grid points tracked at 6-minute intervals during the statistical period is sampled in a 5X5 grid range, and then a vertical profile time series feature dataset of the polarization quantity that represents the precursor of severe convection is generated.

[0075] Definition of actual label: The research scope is divided into 0.05° equally spaced grid points. Then, based on the neighborhood method, if there is an automatic station hourly rainfall >= 20mm / h within a radius of 40km for each grid point, the grid point is defined as a short-term rainstorm label; if there is an automatic station hourly maximum wind >= 8 within a radius of 40km for each grid point, and there is a lightning record within a radius of 25km, the grid point actual situation is a thunderstorm gale label.

[0076] According to the time series characteristic data set of vertical profile of polarization quantity characterizing the precursor of severe convection, the maximum likelihood estimation method is used to fit the time series characteristic variables of vertical profile of polarization quantity with piecewise Gaussian function, and the prior probability corresponding to the time series characteristic variables of vertical profile of polarization quantity within the first hour, the second hour and the third hour before the disaster is constructed. The corresponding benchmark posterior probability (PRO S1 , PRO S2 , PRO S3 ). PRO per hour S It is a 3-dimensional probability matrix, where the x-direction represents the time dimension, the y-direction represents the height dimension, and the z-direction represents the distribution dimension of the effective values ​​of the time series feature variables.

[0077] The method of converting prior to posterior probability includes: since the model is based on the probability threshold method to determine whether a convective disaster occurs, the mean position of the prior statistics of the variables in the model is first used as the reference position (100% probability position), and the prior probabilities of other positions are proportionally stretched and mapped; then, it is simply agreed that the minimum and maximum values ​​of the dual-polarization radar variables in the model correspond to 0% and 200% posterior probabilities, respectively, and the mean position is the 100% probability position; and when the value of the dual-polarization radar variable is greater than the mean, its posterior probability value is 200% probability minus its prior probability value.

[0078] On the basis of the above-mentioned embodiment, in this embodiment, the time series characteristics of each polarization quantity vertical profile and the actual posterior probability of each polarization quantity vertical profile time series characteristics are two-dimensional matrices;

[0079] The horizontal direction of the two-dimensional matrix is ​​the time dimension, and the vertical direction is the height dimension.

[0080] On the basis of the above embodiment, in this embodiment, the actual posterior probability is compared with the probability threshold, and whether severe convection will occur in the future time period is determined according to the comparison result, including:

[0081] Calculate the percentage of actual posterior probabilities greater than the probability threshold in the two-dimensional matrix of actual posterior probabilities of the time series characteristics of the vertical profiles of each polarization quantity;

[0082] When the percentage is greater than a preset threshold, determining that the time series characteristics of the vertical profiles of the polarization quantities have precursor characteristics of severe convection;

[0083] When the time series characteristics of the vertical profiles of all polarization quantities have the precursory characteristics of severe convection, it is determined that severe convection will occur in the future time period.

[0084] For example, calculate the PRO of the time series characteristic variable of the vertical profile of each polarization quantity O The two-dimensional matrix array is greater than the probability threshold PRO TH If the percentage is greater than 70%, it is determined that the current vertical profile time series characteristic variable of the polarization quantity has the precursory characteristics of severe convection in terms of value. If all the vertical profile time series characteristic variables of the polarization quantity have the precursory characteristics of severe convection, it is determined that severe convection disasters will occur in the future period.

[0085] Based on the above embodiment, the step of optimizing the probability threshold using the deep learning model in this embodiment includes:

[0086] Determine the actual posterior probability of the time series feature of the vertical profile of the polarization quantity in the time series feature training set of the vertical profile of the polarization quantity;

[0087] Performing Gaussian fitting on the actual posterior probability of the time series characteristics of the vertical profile of the polarization quantity in the training set to obtain a Gaussian curve;

[0088] Determine the probability difference between the left end of the Gaussian curve and the variance and twice the variance, and determine the search range of the probability threshold according to the two probability differences;

[0089] The actual posterior probability of the time series feature of the vertical profile of the polarization quantity in the training set is calculated through multiple convolution layers along the time dimension to obtain a posterior probability convolution index;

[0090] Utilizing the posterior probability convolution index of each convolution layer in the training set, calculating the optimal risk score corresponding to each convolution layer under the specified probability threshold within the search range;

[0091] The convolution layer corresponding to the maximum value in the optimal risk score is used as the optimal convolution layer, and the probability threshold corresponding to the maximum value in the optimal risk score is used as the optimal probability threshold.

[0092] The deep learning model is mainly composed of two parts: the feature extraction module and the solution module. The feature extraction module is used to calculate the actual posterior probability PRO of the vertical profile time series characteristics of the polarization quantity at the current grid position for the past preset time (1 hour). O Processing, mean filtering and convolution are performed along the time dimension to construct multi-layer convolution layers.

[0093] Different from most deep learning methods that are based on horizontal scale convolution processing, the convolution scheme of the deep learning model designed in this embodiment is based on the violent vertical movement characteristics of strong convection. It performs convolution on the time-series space scale based on the vertical profile direction. Therefore, it retains the original physical and dynamic vertical profile characteristics of the precursor of strong convection to the maximum extent, and at the same time integrates more multi-scale convolution feature information of strong convection evolution in the time dimension.

[0094] Different from the traditional deep learning, which classifies the feature information of multiple convolutional layers according to the weight synthesis strategy, the solution module in this embodiment is used to first calculate the optimal risk score (TS=HitNum / (HitNum+LossNum+FarNum)) corresponding to each convolutional layer based on the training set data using the physical and dynamic probability condition constraint convergence method; select the convolutional layer corresponding to the optimal TS and the convolution index threshold PRO TH As the optimal convolution configuration for this data set. Here HitNum, LossNum and FarNum are the number of hits, the number of misses and the number of empty reports respectively.

[0095] The physical and dynamic probability conditional constraint convergence method assumes that the posterior probability PRO of the time series characteristic variables of the vertical profile that characterizes the physical and dynamic O The probability distribution on the left side of is approximately a semi-Gaussian function characteristic distribution. For example, in the posterior probability PRO O The probability of the position corresponding to the variance σ on the left side of the 100% position (the position where the mean is located) (according to the variance value, the corresponding position is found by step-by-step traversal search) is about 0.6, and the probability of the position corresponding to the variance σ twice the 100% position on the left side is about 0.134. Therefore, the PRO corresponding to the convolutional layer TH The solution of is limited to the range between 0.134 and 0.6. In this range, as PRO TH As the hit rate POD increases, the false alarm rate FAR decreases, and the missed alarm rate LOSS increases. Therefore, in theory, we can always search and find a balance between POD, FAR, and LOSS, so that the TS performance reaches the global optimal level.

[0096] The convergence solution here can be achieved by parallel computing TS at equal intervals, or by taking a binary computing search strategy to calculate the convolutional layer and PRO corresponding to the optimal TS. TH .

[0097] This embodiment is based on the physical and dynamic constraints of the severe convective precursor climate state, and can control the stability of the deep learning training process to the greatest extent, while finally converging to a relatively reasonable optimal result.

[0098] On the basis of the above-mentioned embodiment, the widths of the convolution windows of the multiple convolution layers in this embodiment are different, and the lengths are equal to the maximum height of the convective cloud grid points.

[0099] For example, the deep learning model has 4 convolutional layers, with 10 layers in the x-direction (sampled every 6 minutes for 1 hour) and 30 layers in the y-dimension. O , the convolution windows of the four convolutional layers can be set to 10x30, 5x30, 3x30 and 1x30 respectively.

[0100] On the basis of the above embodiments, the vertical profile time series characteristics of the polarization quantity in the data set described in this embodiment include physical and dynamic vertical profile time series characteristics, and the physical and dynamic vertical profile time series characteristics include one or more of the time series vertical profile of the medium-γ scale horizontal reflectivity factor, the time series vertical profile of the medium-γ scale differential reflectivity, the time series vertical profile of the medium-γ scale differential propagation phase shift rate, the time series vertical profile of the medium-β scale vertical wind shear, the time series vertical profile of the medium-β scale jet stream characteristic, the time series vertical profile of the medium-β scale vorticity characteristic, and the time series vertical profile of the medium-β scale divergence characteristic.

[0101] The meso-γ scale is a horizontal scale between 2km and 20km, and the meso-β scale is a horizontal scale between 20km and 200km. Each time series characteristic sample of severe convection in the data set may include the time series characteristic variables of the physical and dynamic vertical profiles as shown in Table 1:

[0102] Table 1

[0103] Variable Name Full variable name Physical meaning <![CDATA[TSVP ZH ]]> Time series vertical profile of meso-γ scale horizontal reflectivity factor The vertical structure of precipitation intensity representing strong weather characteristics in the atmosphere changes with time based on the three-dimensional grid data sampling of the horizontal reflectivity factor of the contour surface at consecutive times. <![CDATA[TSVP ZDR ]]> Vertical profile of meso-γ-scale differential reflectivity time series The three-dimensional grid data of differential reflectivity of contour surfaces at consecutive times is obtained by sampling, which characterizes the change of the vertical structure of the significant precipitation particles in the atmosphere representing the characteristics of strong weather over time. <![CDATA[TSVP KDP ]]> Time series vertical profile of phase shift rate of meso-γ scale differential propagation The three-dimensional grid data of differential propagation phase shift rate based on the contour surface of continuous time is obtained, which characterizes the change of vertical structure of liquid water content in the atmosphere representing the characteristics of strong weather over time. <![CDATA[TSVP VS ]]> Vertical profile of meso-β scale vertical wind shear time series The vertical wind shear structure of the vertical wind shear in the atmosphere, which represents the characteristics of strong weather, changes over time based on the three-dimensional grid data sampling of the vertical wind shear on the contour surface at consecutive times. <![CDATA[TSVP JET ]]> Vertical profile of meso-β-scale jet characteristics in time series The three-dimensional grid data of wind speed jet streams on the contour surface are obtained based on continuous sampling, which characterizes the change of the vertical structure of the jet stream representing strong weather characteristics in the atmosphere over time. <![CDATA[TSVP VORT ]]> Vertical profile of meso-β scale vorticity characteristics time series The data is obtained based on the three-dimensional grid data sampling of the contour surface vorticity at consecutive times, which characterizes the change of the vertical structure of vorticity in the atmosphere representing the characteristics of strong weather over time. <![CDATA[TSVP DIV ]]> Vertical profile of the time series of meso-β-scale divergence characteristics The three-dimensional grid data of the divergence of the contour surface is obtained based on continuous sampling, which characterizes the change of the vertical structure of the divergence representing the characteristics of strong weather in the atmosphere over time.

[0104] On the basis of the above embodiments, in this embodiment, the polarization amount of the isoheight surface includes one or more of a horizontal reflectivity factor, a differential reflectivity, and a differential propagation phase shift rate;

[0105] The three-dimensional grid data of the contour surface wind field includes one or more of the three-dimensional grid data of the contour surface vertical wind shear, the three-dimensional grid data of the contour surface vorticity, the three-dimensional grid data of the contour surface divergence and the three-dimensional grid data of the contour surface wind speed jet.

[0106] Based on the three-dimensional grid data of the contour surface wind field at preset time intervals within the networking range, the three-dimensional grid data of the contour surface vertical wind shear at preset time intervals can be calculated by subtracting the vector difference of the high-level wind speed from the lowest-level wind speed; the three-dimensional grid data of the contour surface vorticity at preset time intervals can be calculated using the vorticity calculation formula, and the three-dimensional grid data of the contour surface divergence at preset time intervals can be calculated using the divergence calculation formula; the three-dimensional grid data of the contour surface wind speed jet at preset time intervals can be calculated using the low-pass filtering method.

[0107] On the basis of the above embodiments, in this embodiment, before identifying convective clouds according to the polarization amount of the contour surface and the three-dimensional grid data of the wind field at preset time intervals (e.g., every 6 minutes) within the radar networking range in the historical time period, the following is further included:

[0108] Perform quality control on the S-band dual-polarization radar data in the historical time period (e.g., the previous hour) within the radar network range to generate polarization data of a single radar three-dimensional grid point;

[0109] The vertical wind profile data representing the atmospheric environment wind field information in the historical time period within the radar network range is generated by using the velocity-azimuth display (VAD) method;

[0110] The polarization data and vertical wind profile data of all radars within the radar networking range are gridded using a distance exponential weighted jigsaw method to obtain three-dimensional grid data of the polarization and wind field on the contour surface.

[0111] Figure 2 This is a comparison chart of the forecast and actual situation of short-term heavy rainfall (hourly rainfall >= 20mm) in the next two hours at 20:00 in a certain place. Figure 2 (a) is the superposition of the short-term heavy rainfall deterministic forecast result (black curve) predicted at 21:00 using the operational potential forecasting technology and the corresponding hourly actual situation (gray block area); Figure 2 (b) is the superposition of the deterministic forecast result of short-term heavy precipitation at 21:00 using the technology of the present invention (black probability curve) and the corresponding hourly actual situation (gray block area); Figure 2 (c) is the superposition of the short-term heavy rainfall deterministic forecast result (black curve) predicted at 22:00 using the operational potential forecasting technology and the corresponding hourly actual situation (gray block area); Figure 2 (d) in the figure is the superposition of the deterministic forecast result of short-term heavy rainfall at 22:00 predicted by the technology of the present invention (black probability curve) and the corresponding hourly actual situation (grey block area).

[0112] The 0-2 hour deterministic forecast results of dispersed severe convection in a certain place in the afternoon of the 22nd were tested and evaluated. Compared with the deterministic forecast results of short-term heavy precipitation (hourly rainfall>=20mm) of the business potential forecast, it is shown that the new forecast method provided by the present invention can improve the TS score of the deterministic forecast of short-term heavy precipitation in the first hour of the inland area of ​​Zhejiang Province from the original 0.56 to 0.97, and improve the TS score of the deterministic forecast of short-term heavy precipitation in the second hour from the original 0.33 to 0.88. Therefore, this method has a good 0-2 hour advance prediction capability for both systematic and dispersed severe convective processes, and better makes up for the model forecast deviation problem in the short-term near period and the technical defects of the traditional short-term extrapolation that cannot predict the development and strengthening of convective propagation.

[0113] 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 make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A severe convection short-term forecasting method based on deep learning and vertical profile of polarization quantity, characterized in that: include: Convective clouds are identified based on the three-dimensional grid data of polarization and wind field on the contour surface at preset intervals within the radar network range in the historical time period; Tracking the convective clouds using the CLTREC method, establishing a front-to-back matching relationship of the convective clouds, and determining the temporal grid point positions of the convective clouds in the historical time period according to the front-to-back matching relationship; According to the time-series grid point positions of the convective cloud, sampling the time-series characteristics of the vertical profile of the polarization amount of the convective cloud grid points from the three-dimensional grid point data, and determining the actual posterior probability of the time-series characteristics of the vertical profile of the polarization amount of the convective cloud grid points; Comparing the actual posterior probability with a probability threshold, and determining whether severe convection will occur in a future time period according to the comparison result, wherein the probability threshold is optimized using a deep learning model; Determining the actual posterior probability of the vertical profile time series characteristics of the polarization amount of the convective cloud grid point includes: Establish a dataset of time series characteristics of vertical profiles of polarization quantity that characterize the precursors of severe convection; Using the maximum likelihood estimation method to perform piecewise Gaussian function fitting on the time series characteristics of the vertical profile of the polarization quantity in the data set, a priori probability of the occurrence of strong convection within a preset time before the occurrence of the strong convection is constructed; The prior probability is converted by using a priori to posterior probability conversion method to obtain the actual posterior probability of the convective cloud grid point; The step of optimizing the probability threshold using a deep learning model includes: Determine the actual posterior probability of the time series feature of the vertical profile of the polarization quantity in the time series feature training set of the vertical profile of the polarization quantity; Performing Gaussian fitting on the actual posterior probability of the time series characteristics of the vertical profile of the polarization quantity in the training set to obtain a Gaussian curve; Determine the probability difference between the left end of the Gaussian curve and the variance and twice the variance, and determine the search range of the probability threshold according to the two probability differences; The actual posterior probability of the time series feature of the vertical profile of the polarization quantity in the training set is calculated through multiple convolution layers along the time dimension to obtain a posterior probability convolution index; Utilizing the posterior probability convolution index of each convolution layer in the training set, calculating the optimal risk score corresponding to each convolution layer under the specified probability threshold within the search range; The convolution layer corresponding to the maximum value in the optimal risk score is used as the optimal convolution layer, and the probability threshold corresponding to the maximum value in the optimal risk score is used as the optimal probability threshold.

2. The method for short-term severe convection prediction based on deep learning and vertical profile of polarization quantity according to claim 1, characterized in that: The method of tracking the convective cloud by using the CLTREC method and establishing a front-to-back matching relationship of the convective cloud also includes: Perform vertical motion salient feature position correction for each tracking point position, correct the position P VM The formula is as follows: ； ； ； in, Indicates that the initial position ( X , Y ) as the center, and search the differential propagation phase shift rate based on the grid point within the preset radius area. K DP Calculated vertical liquid water content The position of the maximum value, the initial position is the grid point corresponding to the current convective cloud, cappiNum is the highest layer of the grid point, is the moving characteristics of convective clouds at adjacent moments The rate of change.

3. The method for short-term severe convection prediction based on deep learning and vertical profile of polarization quantity according to claim 1, characterized in that: The time series characteristics of the vertical profiles of each polarization quantity and the actual posterior probability of the time series characteristics of the vertical profiles of each polarization quantity are two-dimensional matrices; The horizontal direction of the two-dimensional matrix is ​​the time dimension, and the vertical direction is the height dimension.

4. The method for short-term severe convection prediction based on deep learning and vertical profile of polarization quantity according to claim 3, characterized in that: Comparing the actual posterior probability with the probability threshold, and determining whether severe convection will occur in a future time period according to the comparison result, including: Calculate the percentage of actual posterior probabilities greater than the probability threshold in the two-dimensional matrix of actual posterior probabilities of the time series characteristics of the vertical profiles of each polarization quantity; When the percentage is greater than a preset threshold, determining that the time series characteristics of the vertical profiles of each polarization quantity have precursor characteristics of severe convection; When the time series characteristics of the vertical profiles of all polarization quantities have the precursory characteristics of severe convection, it is determined that severe convection will occur in the future time period.

5. The method for short-term severe convection prediction based on deep learning and vertical profile of polarization quantity according to claim 1, characterized in that: The convolution windows of the multiple convolution layers have different widths, and their lengths are equal to the maximum height of the convective cloud grid points.

6. The method for short-term severe convection prediction based on deep learning and vertical profile of polarization quantity according to claim 1, characterized in that: The vertical profile time series characteristics of the polarization quantity in the data set include physical and dynamic vertical profile time series characteristics, and the physical and dynamic vertical profile time series characteristics include one or more of the time series vertical profile of the meso-γ scale horizontal reflectivity factor, the time series vertical profile of the meso-γ scale differential reflectivity, the time series vertical profile of the meso-γ scale differential propagation phase shift rate, the time series vertical profile of the meso-β scale vertical wind shear, the time series vertical profile of the meso-β scale jet stream characteristic, the time series vertical profile of the meso-β scale vorticity characteristic, and the time series vertical profile of the meso-β scale divergence characteristic.

7. The method for short-term severe convection prediction based on deep learning and vertical profile of polarization quantity according to any one of claims 1 to 6, characterized in that: The polarization quantity of the contour plane includes one or more of a horizontal reflectivity factor, a differential reflectivity, and a differential propagation phase shift rate; The three-dimensional grid data of the contour surface wind field includes one or more of the three-dimensional grid data of the contour surface vertical wind shear, the three-dimensional grid data of the contour surface vorticity, the three-dimensional grid data of the contour surface divergence and the three-dimensional grid data of the contour surface wind speed jet.

8. The method for short-term severe convection prediction based on deep learning and vertical profile of polarization quantity according to any one of claims 1 to 6, characterized in that: Before identifying convective clouds based on the three-dimensional grid data of polarization quantities and wind fields of contour surfaces at preset intervals within the radar network range in the historical time period, it also includes: The S-band dual-polarization radar data in the historical time period within the radar network is quality controlled to generate polarization data of a single radar three-dimensional grid point; Generate vertical wind profile data representing atmospheric environment wind field information within a historical period within the radar network range by using a speed and azimuth display method; The polarization data and vertical wind profile data of all radars within the radar networking range are gridded using a distance exponential weighted jigsaw method to obtain three-dimensional grid data of the polarization and wind field on the contour surface.

Images