Meteorological convection simulation method and system based on machine learning

By coupling the WRF model with the machine learning model to calculate the convection tendency and precipitation rate, the problem that numerical weather forecast model is difficult to characterize convection within the atmospheric boundary layer is solved, and the accuracy and effect of meteorological prediction are improved.

CN119089781APending Publication Date: 2024-12-06SHENZHEN INST OF ARTIFICIAL INTELLIGENCE & ROBOTICS FOR SOC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411180018.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-26
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

The existing numerical weather forecast model is difficult to accurately characterize convection within the atmospheric boundary layer, resulting in poor meteorological prediction effects, especially in warm rainy areas.

Method used

The meteorological convection simulation method based on machine learning is used to couple the WRF model with the machine learning model. By calculating the cloud relaxation time counter NCA, the characteristic data of the convection cloud meets the convection occurrence conditions, and calculate the convection tendency based on the bidirectional long and short-term memory model. and precipitation rate, updated average vertical velocity to improve meteorological prediction accuracy.

Benefits of technology

It improves the accuracy and effectiveness of meteorological prediction, can better handle multi-scale and gray area convection, and enhances the simulation ability of the convection process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119089781A_ABST
    Figure CN119089781A_ABST
Patent Text Reader

Abstract

The invention provides a meteorological convection simulation method and system based on machine learning, and the method comprises the steps: inputting grid parameters when a WRF model is integrated to a t moment, and carrying out the calculation to obtain an NCA; if the NCA is smaller than the time threshold value, feature data of the streaming cloud are calculated; judging whether the feature data of the convective cloud meets a convective occurrence condition or not; when the convection occurrence condition is met, calculating the convection tendency and the precipitation rate based on a specified machine learning model, updating the average vertical speed, and executing specified steps; when the convection occurrence condition is not met, the average vertical speed is updated, the convection tendency and the precipitation rate are set to be 0, and specified steps are executed; the specified step comprises the steps that NCA is reduced by 1, t is increased by 1, and the step of inputting the grid parameters when the WRF model is integrated to the t moment is executed again. According to the scheme, the WRF model and the machine learning model are coupled, and machine learning is applied to weather convection simulation to accurately predict weather, so that the weather prediction effect is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a meteorological convection simulation method and system based on machine learning. Background Art

[0002] Heavy rainfall in warm areas is regional and long-lasting, which poses a significant challenge to accurate weather forecasting. Currently, numerical weather forecast models are usually used to predict heavy rainfall in warm areas.

[0003] However, as the grid spacing of existing numerical weather forecast models approaches the scale of the convective process, existing numerical weather forecast models find it difficult to characterize the convection within the atmospheric boundary layer, which leads to the inability to accurately predict the weather and poor weather forecast results. Summary of the invention

[0004] In view of this, an embodiment of the present invention provides a meteorological convection simulation method and system based on machine learning to solve the problems of poor meteorological forecasting effect existing in existing numerical weather forecast models.

[0005] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0006] A first aspect of an embodiment of the present invention discloses a meteorological convection simulation method based on machine learning, the method comprising:

[0007] The grid parameters are input into the Weather Research and Forecasting WRF model when it is integrated to time t to calculate the cloud relaxation time counter NCA;

[0008] If the NCA is less than the time threshold, calculating characteristic data of the convective cloud;

[0009] Determining whether the characteristic data of the convective cloud meets the convection occurrence condition;

[0010] When the convection occurrence condition is met, the convection tendency and precipitation rate are calculated based on the specified machine learning model, and the average vertical velocity is updated, and the specified steps are executed; wherein the convection tendency and the precipitation rate are used to perform meteorological convection simulation in the WRF model, and the average vertical velocity is used to maintain the consistency of physical quantities;

[0011] When the convection occurrence condition is not met, updating the average vertical speed, setting the convection tendency and the precipitation rate to 0, and executing the specified step;

[0012] The designation step comprises:

[0013] The NCA is reduced by 1 and t is increased by 1, and the step of inputting the grid parameters when the Weather Research and Forecasting WRF model is integrated to time t is returned to be executed until the WRF model completes the iteration.

[0014] Preferably, if the NCA is less than a time threshold, calculating characteristic data of convective clouds includes:

[0015] If the NCA is less than the time threshold, a one-dimensional cloud model is used based on the multi-scale Kain-Fritsch algorithm, and the characteristic data of the convective cloud are calculated in combination with the temperature, dew point temperature, air pressure, wind speed and wind direction. The characteristic data of the convective cloud include at least: lifting condensation height, convective effective energy potential, average vertical velocity, cloud top height, cloud base height and entrainment rate.

[0016] Preferably, determining whether the characteristic data of the convective cloud meets the convection occurrence condition includes:

[0017] Determining whether the convection effective energy level exceeds a convection effective energy level threshold;

[0018] If the convection effective energy level exceeds the convection effective energy level threshold, it is determined that the convection occurrence condition is met;

[0019] If the convection effective energy level does not exceed the convection effective energy level threshold, it is determined that the convection occurrence condition is not met.

[0020] Preferably, when the convection occurrence condition is met, the convection tendency and precipitation rate are calculated based on a specified machine learning model, and the average vertical velocity is updated, and the specified steps are performed, including:

[0021] When the convection occurrence condition is met, at the grid point where the convection occurrence condition has been met, a multi-scale Kain-Fritsch algorithm based on a specified machine learning model is used to calculate the convection tendency and precipitation rate, wherein the specified machine learning model is a bidirectional long short-term memory model;

[0022] Updates the average vertical velocity, taking the specified number of steps.

[0023] Preferably, after calculating the cloud relaxation time counter NCA, the method further includes:

[0024] If the NCA is greater than or equal to the time threshold, the average vertical velocity is updated, the convective tendency and the precipitation rate remain unchanged, and the specified step is performed.

[0025] A second aspect of an embodiment of the present invention discloses a meteorological convection simulation system based on machine learning, the system comprising:

[0026] The first calculation unit is used to input the grid parameters when the Weather Research and Forecasting WRF model is integrated to time t to calculate the cloud relaxation time counter NCA;

[0027] A second calculation unit, configured to calculate characteristic data of convective clouds if the NCA is less than a time threshold;

[0028] A judging unit, used to judge whether the characteristic data of the convective cloud meets the convection occurrence condition;

[0029] a third calculation unit, configured to calculate the convection tendency and precipitation rate based on a specified machine learning model and update the average vertical velocity to execute a specified step when the convection occurrence condition is met; wherein the convection tendency and precipitation rate are used to perform meteorological convection simulation in the WRF model, and the average vertical velocity is used to maintain the consistency of physical quantities;

[0030] A first updating unit, configured to update the average vertical velocity, set the convection tendency and the precipitation rate to 0, and execute the specified step when the convection occurrence condition is not met;

[0031] The designation step comprises:

[0032] The NCA decreases by 1 and t increases by 1, and the first calculation unit is executed again until the WRF model completes the iteration.

[0033] Preferably, the second calculation unit is specifically used to: if the NCA is less than a time threshold, use a one-dimensional cloud model based on a multi-scale Kain-Fritsch algorithm, combine temperature, dew point temperature, air pressure, wind speed and wind direction to calculate characteristic data of convective clouds, wherein the characteristic data of convective clouds include at least: lifting condensation height, convective effective energy potential, average vertical velocity, cloud top height, cloud bottom height and entrainment rate.

[0034] Preferably, the judgment unit is specifically used to: judge whether the convection effective energy level exceeds the convection effective energy level threshold; if the convection effective energy level exceeds the convection effective energy level threshold, determine that the convection occurrence condition is met; if the convection effective energy level does not exceed the convection effective energy level threshold, determine that the convection occurrence condition is not met.

[0035] Preferably, the third calculation unit is specifically used for: when the convection occurrence condition is met, at the grid point where the convection occurrence condition has been met, using a multi-scale Kain-Fritsch algorithm based on a specified machine learning model to calculate the convection tendency and precipitation rate, wherein the specified machine learning model is a bidirectional long short-term memory model; updating the average vertical velocity and executing the specified steps.

[0036] Preferably, it also includes:

[0037] The second updating unit is used to update the average vertical speed if the NCA is greater than or equal to the time threshold, keep the convective tendency and the precipitation rate unchanged, and execute the specified step.

[0038] Based on the above-mentioned embodiment of the present invention, a method and system for simulating meteorological convection based on machine learning is provided. The method is as follows: when the WRF model is integrated to time t, grid parameters are input to calculate NCA; if NCA is less than the time threshold, characteristic data of convective clouds are calculated; it is determined whether the characteristic data of convective clouds meet the conditions for convection occurrence; when the conditions for convection occurrence are met, the convection tendency and precipitation rate are calculated based on the specified machine learning model, and the average vertical velocity is updated, and the specified steps are executed; when the conditions for convection occurrence are not met, the average vertical velocity is updated, and the convection tendency and precipitation rate are set to 0, and the specified steps are executed; the specified steps include: NCA is reduced by 1 and t is increased by 1, and the step of inputting grid parameters when the WRF model is integrated to time t is returned to execute until the WRF model completes the iteration. This scheme couples the WRF model and the machine learning model, and applies machine learning to meteorological convection simulation to accurately predict the weather, thereby improving the effect of meteorological prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0040] Figure 1 A flow chart of a meteorological convection simulation method based on machine learning provided in an embodiment of the present invention;

[0041] Figure 2 A structural example diagram of a Bi-LSTM model provided in an embodiment of the present invention;

[0042] Figure 3 Another flow chart of a meteorological convection simulation method based on machine learning provided by an embodiment of the present invention;

[0043] Figure 4 An example diagram comparing the results of offline data processing between the traditional MSKF technology provided in an embodiment of the present invention and this solution;

[0044] Figure 5 An example diagram of a vertical cross-section of convective tendency related error statistics provided by an embodiment of the present invention;

[0045] Figure 6 First example plot of the mean absolute difference between the average spatial predictions generated for the original WRF model and the WRF-ML model simulations;

[0046] Figure 7Second example plot of the mean absolute difference between the average spatial predictions generated by the original WRF model and the WRF-ML model simulations;

[0047] Figure 8 A third example plot of the mean absolute difference between the mean spatial predictions generated for the original WRF model and the WRF-ML model simulations;

[0048] Fig. 9 An example diagram of the comparison of domain average time series between the original WRF simulation and the WRF-ML simulation provided by the embodiment of the present invention;

[0049] Fig.10 A structural block diagram of a meteorological convection simulation system based on machine learning provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 are within the scope of protection of the present invention.

[0051] In this application, the terms "comprises", "comprising" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or device comprising the element.

[0052] In order to better understand the following content, some terms are explained here:

[0053] Bi-LSTM (bidirectional long short-term memory): Bidirectional long short-term memory model, also known as Bi-LSTM model;

[0054] WRF (Weather Research and Forecast): Weather Research and Forecast, also known as WRF model or WRF mode; WRF model is a numerical weather forecast and climate simulation tool. WRF model is highly configurable and modular, supporting different physical modes to meet various application requirements. This solution can better simulate severe weather processes by using WRF model for convective forecasting.

[0055] MSKF (multi-scale Kain-Fritsch): multi-scale Kain-Fritsch convection parameterization scheme, also known as multi-scale Kain-Fritsch algorithm or multi-scale Kain-Fritsch convection parameterization scheme;

[0056] WRF-ML coupler (Weather Research and Forecast and machine learning coupler): WRF and machine learning model coupling framework.

[0057] Warm-zone rainstorms are regional and long-lasting, which poses a significant challenge to accurate weather forecasting. Although the rainfall in warm zones is small, sudden and extreme precipitation events can cause losses such as flooding houses and destroying farmland. Existing studies have shown that the performance of convective rainfall forecasts can be improved by accurately analyzing terrain features and higher spatial resolution. Currently, scientific research institutions and other entities have recognized the importance of resolution in forecasting severe convective weather, and have begun to attach importance to high-resolution operational forecasting models for warm-zone rainstorms and sudden local rainstorms, so as to more accurately characterize terrain effects.

[0058] Due to the complex coupling of various meteorological factors, such as low-level jets, land-sea contrast, topography and urban buildings, complex topography and heterogeneous land surfaces in regions such as South China play a key role in promoting active convection. Currently, numerical weather forecast models are commonly used to predict heavy rainfall in warm areas.

[0059] However, as the grid spacing of existing numerical weather forecast models approaches the scale of the convective process, existing numerical weather forecast models find it difficult to characterize the convection within the atmospheric boundary layer, which leads to the inability to accurately predict the weather and poor weather forecast results.

[0060] Further research found that when the traditional Kain-Fritsch (KF) convection analysis technology is used to deal with gray zone (scale between 1-10 kilometers) convection and multi-scale convection, the triggering mechanism and parameter analysis method of the traditional KF convection analysis technology are insufficient, making it difficult to accurately reproduce these complex convection processes.

[0061] The multi-scale Kain-Fritsch (MSKF) convection analysis technology is still based on empirical formulas, lacks flexibility and adaptability, and is difficult to fully cover the complex convection processes under different weather conditions.

[0062] When machine learning algorithms are applied to convection parameter analysis, machine learning models require a large amount of training data and computing resources. The interpretability of machine learning models is low and there may be stability issues during online synthesis.

[0063] In response to the above situation, this scheme introduces machine learning methods into convection analysis and couples the WRF model with the machine learning model, which can better handle multi-scale convection and gray zone convection and improve meteorological forecasting results.

[0064] See also Figure 1 , shows a flow chart of a meteorological convection simulation method based on machine learning provided by an embodiment of the present invention, the meteorological convection simulation method comprising:

[0065] Step S101: When the WRF model is integrated to time t, grid parameters are input to calculate NCA.

[0066] It should be noted that NCA (counter of the cloud relaxation time) is a cloud relaxation time counter.

[0067] In the specific implementation of step S101, when the WRF model is integrated to time t, grid parameters are input to calculate the variable NCA, wherein the grid parameters include but are not limited to temperature, humidity, pressure, wind speed, etc. The specific meaning of the grid parameters can be found in Table 1 shown later.

[0068] Step S102: Determine whether NCA is less than a time threshold. If NCA is greater than or equal to the time threshold, execute step S103; if NCA is less than the time threshold, execute step S104.

[0069] In the specific implementation of step S102, after calculating NCA, it is determined whether NCA is less than the time threshold; if NCA is greater than or equal to the time threshold, step S103 is executed; if NCA is less than the time threshold, step S104 is executed.

[0070] In a specific embodiment, the time threshold mentioned above may be half of the time step (denoted as dt) during the integration of the WRF model;

[0071] That is, it is determined whether NCA is less than 0.5dt; if NCA is less than 0.5dt, step S104 is executed;

[0072] If NCA is greater than or equal to 0.5dt, it indicates that convection is in progress, and step S103 is executed without updating the convection tendency and precipitation rate.

[0073] Step S103: Update the average vertical velocity, keep the convection tendency and precipitation rate unchanged, and execute the designated step of step S108.

[0074] In the specific implementation of step S103, if NCA is greater than or equal to the time threshold (0.5dt), the average vertical velocity (denoted as w0avg) is updated, the convective tendency and precipitation rate remain unchanged, and the designated step S108 is executed, wherein step S108 is the "designated step" mentioned in this solution.

[0075] Step S104: Calculate characteristic data of convective clouds.

[0076] In the specific implementation of step S104, if NCA is less than the time threshold (0.5dt), a one-dimensional cloud model is used based on a multi-scale Kain-Fritsch algorithm (MSKF algorithm), and the temperature, dew point temperature (or humidity), air pressure, wind speed and wind direction are combined to calculate the characteristic data of the convective cloud. The characteristic data of the convective cloud at least include: lifting condensation level (LCL), convective available potential energy (CAPE), average vertical velocity (w0avg), cloud top height, cloud bottom height and entrainment rate, etc.

[0077] It should be noted that the characteristic data of convective clouds can be used to assess the possibility of convection generation, among which LCL is crucial to determine the occurrence of potential convective activities, and lower LCL is conducive to more intense convection;

[0078] CAPE quantifies the buoyancy capacity required for an air parcel to form deep convective clouds when it reaches the level of freeconvection (LFC) above its LCL. Higher CAPE values ​​indicate more intense potential convection.

[0079] The cloud base is usually located at the LCL, while the cloud top is defined as the altitude where buoyancy becomes negligible, and the vertical extent between the cloud base and the cloud top affects the cloud growth and precipitation potential.

[0080] Step S105: Determine whether the characteristic data of the convective cloud meets the convection occurrence condition. If the convection occurrence condition is met, execute step S106; if the convection occurrence condition is not met, execute step S107.

[0081] In the process of implementing step S105, it is determined whether the characteristic data of the convective cloud meets the convection occurrence conditions (that is, it is checked whether the CAPE and the cloud thickness meet the convection occurrence conditions); when the convection occurrence conditions are met, step S106 is executed; when the convection occurrence conditions are not met, step S107 is executed.

[0082] In some embodiments, the specific method for determining whether the convection occurrence condition is met is: determining whether the convection effective energy level (CAPE) exceeds the convection effective energy level threshold (called the CAPE threshold); if the convection effective energy level exceeds the convection effective energy level threshold, determining that the convection occurrence condition is met; if the convection effective energy level does not exceed the convection effective energy level threshold, determining that the convection occurrence condition is not met.

[0083] In other embodiments, the tropospheric low cloud cover (CLD) can also be used to determine whether the convection occurrence conditions are met, that is, to determine whether the tropospheric low cloud cover (CLD) is exceeded; if the tropospheric low cloud cover (CLD) is exceeded, the convection occurrence conditions are met, otherwise the convection occurrence conditions are not met.

[0084] It should be noted that by determining whether the CAPE threshold, the tropospheric low cloud cover, and evaluating the entrainment rate, the impact of ambient air on the evolution of the convective system can be measured.

[0085] Step S106: Calculate the convection tendency and precipitation rate based on the specified machine learning model, and update the average vertical speed, and execute the specified step S108.

[0086] In the process of implementing step S106, when the convection occurrence conditions are met, at the grid points where the convection occurrence conditions have been met (or, at the grid points that trigger convection), a multi-scale Kain-Fritsch algorithm based on a specified machine learning model is used to calculate the convection tendency (including NCA, tendencies, etc.) and precipitation rate (denoted as pratec), wherein the specified machine learning model is a bidirectional long short-term memory model (Bi-LSTM model); the average vertical velocity is updated, and the specified step of step S108 is executed.

[0087] It should be noted that the convective tendency and precipitation rate are used to simulate meteorological convection in the WRF model, and the average vertical velocity is used to maintain the consistency of physical quantities.

[0088] More specifically, the convective tendency can be expressed as the rate of change of temperature, humidity and wind speed during the convection process, and the precipitation rate is obtained by calculating the amount of condensed water during the convection process; the convective tendency and precipitation rate can be used in the WRF model for meteorological forecasts such as temperature, water vapor, cloud water, rain, cloud ice, snow and precipitation.

[0089] It is worth noting that, from the above content, it can be seen that this solution adopts a multi-scale Kain-Fritsch algorithm based on the Bi-LSTM model (specified machine learning model). The structure of the Bi-LSTM model is as follows: Figure 2 As shown in the figure, the Bi-LSTM model is able to perform regression and classification predictions simultaneously.

[0090] like Figure 2As shown in the structure of the Bi-LSTM model, this solution uses a batch processing algorithm to replace the traversal summation algorithm of the traditional MSKF. Figure 2 B in represents the number of processing batches. Each batch of data contains 27 (just for example) features and spans 44 vertical layers. These 22 features are shown in Table 1 below.

[0091] Table 1:

[0092]

[0093]

[0094] It should be noted that the Where p is the air pressure, p0 is the standard air pressure of 1000 hPa; R d is the dry air constant, which is 287 J / (kg.Kelvin); c p is the specific heat capacity of dry air, which is 1.005 kJ / (kg.Kelvin).

[0095] Combination Figure 2 As can be seen from Table 1 above, the dimension of the input data of the Bi-LSTM model is B×27×44. Before the input data is input into the Bi-LSTM model, the input data is preprocessed by a preprocessing module, which includes a one-dimensional convolutional layer. The preprocessing module expands the feature dimension from 27 to 64.

[0096] It should be noted that the variables of type "input and output (bidirectional variables)" in Table 1 above are both input variables and output variables. The variables of type "input and output (bidirectional variables)" will be updated before and after calling the Bi-LSTM model, and the previous output will be used as the next input.

[0097] The input variables and output variables used by the Bi-LSTM model are all in Table 1 above. The output results of the Bi-LSTM model include variables of type "input and output (bidirectional variables)" and "output"; the convection tendency refers to the variables rthcuten, rqvcuten, rqccuten, rqrcuten, rqicuten, rqscuten and NCA in the output results shown in Table 1 above.

[0098] It is worth noting that the output variable raincv represents the time step precipitation caused by convection, which is calculated by multiplying the precipitation rate by the time step of the model. Among all the variables shown in Table 1 above, 5 are two-dimensional surface variables and the rest are three-dimensional variables with 44 vertical sections.

[0099] In addition, the specified machine learning model (Bi-LSTM model) used in the present invention includes four derived variables as inputs, which include: a two-dimensional Boolean variable indicating convection triggering based on the NCA value, the pressure difference between adjacent vertical layers, the saturated water vapor mixing ratio, and the relative humidity.

[0100] Combination Figure 2 As shown in the structural example diagram of the Bi-LSTM model, the model proposed in this scheme includes: a shared Bi-LSTM layer for learning features, a classification subnetwork and a regression subnetwork;

[0101] The shared Bi-LSTM layer consists of 3 repeated Bi-LSTM blocks (bidirectional long short-term memory modules), each of which contains a forward layer and a backward layer with a feature dimension of 32;

[0102] The classification subnetwork consists of a 1×1 one-dimensional convolutional layer, a fully connected layer, and a sigmoid activation layer. The output of the sigmoid activation layer represents the probability distribution of convection triggering. The binary cross entropy loss function is used as the cost function of the classification task;

[0103] The regression subnetwork contains a fully connected layer to output precipitation rate, NCA, and convective tendency;

[0104] Finally, the outputs of the two sub-networks are processed through a post-processing module, which is used to process the variables output by the regression branch to ensure their physical consistency.

[0105] It should be noted that the post-processing module aims to ensure the physical consistency of all variables. To achieve this goal, this solution applies the following rules:

[0106] (1) At the grid points where the input NCA is greater than or equal to 0.5dt (equivalent to when NCA≧0.5dt, convection occurs), all other variables remain unchanged because all other variables are still within the convection life cycle;

[0107] (2) The output NCA must be an integer;

[0108] (3) When predicting grid points where convection is not active, all corresponding output variables are set to 0 by default.

[0109] The physical consistency of all variables is ensured by the above rules (1)-(3).

[0110] As for model training, the Bi-LSTM model provided by this solution includes classification and regression tasks, so the performance of the Bi-LSTM model is optimized by minimizing the multi-task loss function; the loss function is defined as: the sum of the weighted combination of the binary cross entropy loss triggered by the convection and the L1 loss of all output variables in the regression subnetwork; the specific content of the loss function is shown in the following formula (1).

[0111]

[0112] In formula (1), i and j represent the grid points in the domain, p i,j represents the probability of convection being triggered. If convection is triggered, p i,j The true value label is 1, if the convection is not triggered, then p i,j The true value label is 0; the classification loss L cls is calculated using binary cross entropy loss;

[0113] For the regression loss of different variables c, λ c as weights to balance the output variables, taking into account their respective magnitudes; is the probability of convection occurring in real conditions. If convection does not occur, The value is 0. If convection occurs, The value of is 1;

[0114] L1 c The loss function value of the L1 regression branch is calculated on the grid where the convection is actually generated, that is, only Calculated on a grid with a value of 1, if If it is 0, the loss function of the regression branch is not calculated;

[0115] Indicates that the L1 regression loss is only at the triggered grid point Activate, otherwise Disable, both through N cls and N reg Normalize, N cls and N reg They correspond to the total number of grid points and the number of trigger grid points respectively.

[0116] The overall model performance indicators of the Bi-LSTM model include the root mean squared error (RMSE) and the correlation coefficient. The calculation methods of the mean absolute error (MAE) and mean bias error (MBE) of each vertical layer are shown in formulas (2) and (3).

[0117]

[0118] In formula (2) and formula (3), Y(i,l) and Y ML (i, l) represent the outputs of the traditional MSKF technique and the MSKF algorithm based on the specified machine learning model (multi-scale Kain-Fritsch algorithm based on the specified machine learning model) provided by this scheme, i represents the horizontal grid point of the vertical profile, N is the number of horizontal grid points in the domain, and l represents the vertical layer index.

[0119] Step S107: Update the average vertical velocity, set the convection tendency and precipitation rate to 0, and execute the designated step of step S108.

[0120] In the specific implementation of step S107, when the convection occurrence condition is not met, the average vertical speed is updated, and the convection tendency and precipitation rate are set to 0, and the designated step of step S108 is executed.

[0121] Specifically, if the convection occurrence condition is not met (or if the convection condition is not reached), the average vertical velocity is updated, all output variables except the average vertical velocity (w0avg) in the above Table 1 are set to 0, and then the specified step S108 is executed.

[0122] From the contents of the above steps 102 to S107, it can be seen that regardless of the convection state, the average vertical velocity (w0avg) is updated. The average vertical velocity is used to maintain the consistency of physical variables. The average vertical velocity (w0avg) is a variable used to determine whether the grid meets the convection condition.

[0123] Step S108: NCA decreases by 1 and t increases by 1, and the process returns to step S101 until the WRF model completes the iteration.

[0124] In the specific implementation of step S108, NCA is reduced by 1 (equivalent to NCA reducing one model time step), t is increased by 1 (t=t+1, i.e., integrated to the next step), and then returns to execute step S101 until the WRF model completes the iteration.

[0125] That is, in each iteration of the WRF model loop, active convection will cause NCA to reduce one model time step, set t=t+1, and then return to execute step S101.

[0126] It should be noted that, by iterating the above steps S101 to S108 and obtaining a plurality of output results as shown in the above Table 1, the simulation of meteorological convection can be realized.

[0127] In an embodiment of the present invention, the WRF model and the machine learning model are coupled, and machine learning is applied to meteorological convection simulation to accurately predict the weather, thereby improving the effect of meteorological prediction.

[0128] In order to better explain this scheme, Figure 3 Another flowchart of a meteorological convection simulation method based on machine learning is shown as an example. Figure 3 The following steps are involved:

[0129] Step S301: The WRF model is integrated to time t.

[0130] Step S302: Input grid parameters.

[0131] Step S303: Determine whether NCA is less than 0.5dt. If NCA is greater than or equal to 0.5dt, it indicates that convection is in progress, and step S304 is executed; if NCA is less than 0.5dt, step S305 is executed.

[0132] Step S304: Update the average vertical velocity, keep the convective tendency and precipitation rate unchanged, and execute step S309.

[0133] Step S305: Calculate characteristic data of convective clouds.

[0134] Step S306: Check whether CAPE and cloud thickness meet the convection occurrence conditions. If yes, execute step S307; if not, execute step S308.

[0135] Step S307: Calculate the convection tendency and precipitation rate, trigger convection, update w0avg, and feed back the convection tendency to the main program, and execute step S309.

[0136] Step S308: If no convection occurs, w0avg is updated, the convection tendency and precipitation rate are set to 0, and step S309 is executed.

[0137] Step S309: calling other physical parameter optimization solutions.

[0138] Step S310: t=t+1, NCA decreases by 1, and the process returns to step S301.

[0139] It should be noted that the execution principle of step S301 to step S310 can be found in the above embodiment of the present invention. Figure 1 The contents in will not be repeated here.

[0140] In order to better explain the beneficial effects of this solution, compared with the prior art, the following content is used to illustrate the beneficial effects of this solution, which are explained from two aspects: offline data analysis results and real-time analysis.

[0141] (I) Offline data analysis results:

[0142] Combination Figure 4 The figure shows the comparison of the results of offline data processing between the traditional MSKF technology and this scheme. The traditional MSKF technology and this scheme are used to predict NCA, pratec and convection tendency respectively. For the convenience of comparison, the units of precipitation rate and temperature tendency are changed from mms -1 and Ks -1 Convert to mmd -1 and Kd -1 , and a conversion factor of 86400 (24×3600) is applied; similarly, due to convection, the parameters rqvcuten, rqccuten, and rqrcuten are multiplied by 86400000 (24×3600×1000) to convert their dimensions from kgkg -1 s -1 Convert to gkg -1 d -1 ; Therefore, in offline verification, w0avg is calculated using the true value as input, so w0avg is not evaluated in offline evaluation.

[0143] It should be noted that Figure 4 The y-axis is the predicted value and the x-axis is the true value.

[0144] exist Figure 4 Among all the variables shown in , variable NCA shows a significantly higher RMSE value of 4.32, and the data points are widely scattered among a large range of values, which shows that accurately predicting convection is a greater challenge;

[0145] To eliminate dependence on the time step, NCA was divided by the model time step of 15 seconds before plotting and statistical evaluation; precipitation rate shows the highest correlation coefficient and the smallest variability, because most data points are tightly clustered on the "1:1 line"; although temperature and the four humidity tendencies show some degree of variability, most data sets are closely aligned with the "1:1 line", with a correlation coefficient of 0.91 for convective triggering.

[0146] Therefore, this scheme showed a strong correlation with the traditional MSKF technology in all the variables examined, and the correlation coefficient was always higher than 0.91. This shows that this scheme can completely replace the traditional MSKF technology.

[0147] like Figure 5 An example of a vertical profile of the convective tendency related error statistics based on offline data is shown in FIG. Figure 5The solid and dashed lines in the figure represent the MAE and MBE of the inclination at each vertical layer, respectively. In addition, the shaded areas correspond to the 5th and 95th percentiles of the differences in the inclinations predicted by the proposed scheme and the conventional MSKF technique, respectively. The vertical error distributions for all inclinations show a remarkable uniformity, with higher variability observed in the pressure layers between 800 and 1000 hPa, which correspond to the atmospheric layers where convection occurs most frequently.

[0148] Due to the significantly lower abundance of clouds and rain compared to atmospheric water vapor, the margins of error for rqccuten and rqrcuten are significantly lower than that for rqvcuten.

[0149] (II) Real-time analysis:

[0150] This scheme replaces the traditional MSKF with the MSKF algorithm based on the specified machine learning model (multi-scale Kain-Fritsch algorithm based on the specified machine learning model) and combines it with the WRF model to simulate the convection process; this scheme couples WRF with the machine learning algorithm, which is equivalent to the WRF-ML model.

[0151] This scheme simulates the original MRF model and the WRF-ML model provided by this scheme for 168 hours. The aforementioned simulation is performed independently of the training dataset to ensure the generalization ability of the evaluation technology.

[0152] The mean absolute difference (MAD) between the average spatial predictions generated by the original WRF model and the simulations by the WRF-ML model is shown in Figures 6 to 8 Example plot showing the mean absolute difference between the average spatial predictions generated by the original WRF model and the WRF-ML model simulations. Figures 6 to 8 This completes the example graph of this mean absolute difference.

[0153] It should be noted that Figures 6 to 8 The regional divisions in the text are used only to distinguish between land and sea.

[0154] Combination Figures 6 to 8 The forecast results include the accumulated convective precipitation (RAINC) and non-convective precipitation (RAINNC) within 12 hours, and the 2-meter temperature (T2M) at 24, 72, 120 and 168 hours;

[0155] In the spatial predictions, red and blue patterns indicate the magnitude of the predicted values, while in the spatial differences, these colors indicate positive and negative deviations from the ML-based simulations, respectively; the green pattern indicates the smallest deviation from the original WRF simulation.

[0156] In addition, the difference in domain average MAD between the original WRF model and the WRF-ML model in this scheme is small, which shows that there is good consistency between the WRF-ML simulation in this scheme and the original WRF simulation, and the difference in domain average MAD does not increase with the passage of simulation time. For example, when the difference in domain average MAD during 168-hour prediction is compared with that during 24-hour prediction, the difference in domain average MAD does not increase, which indicates that the present invention achieves stable forecast simulation.

[0157] Fig. 9 An example plot showing the domain-averaged time series comparison between the original WRF simulation and the proposed WRF-ML simulation is shown in Figure 1. Fig. 9 Includes 6-hour accumulations of RAINC and RAINNC as well as T2M forecasts.

[0158] from Fig. 9 It can be seen from the content that the prediction effect of this scheme coupled with WRF-ML simulation is better, especially in capturing the diurnal changes of RAINC, RAINNC and T2M, and the prediction performance of T2M is better.

[0159] In practical applications, in order to verify the effect of this solution, the following content is used to illustrate one of the implementation methods of this solution:

[0160] Training and test data were generated based on the WRF model at a horizontal resolution of 5 km. The model settings include 44,000 grid points, a horizontal grid spacing of 5 km, and a grid size of 220×200. The model includes 45 vertical layers with a top air pressure of 50 hPa. The physical and chemical parameters in the model include: WSM-6 micrometeorology, Bougeault-Lacarrère (BouLac), Monin-Obukhov (Janjic) surface layer, UnifiedNoah surface model, RRTMG shortwave and longwave radiation, MSKF cumulus clouds; the simulation step size is set to 15 seconds.

[0161] The data initial and boundary conditions are derived from a designated dataset provided by a designated weather forecast center, which has a horizontal resolution of 0.25° and contains 29 pressure layers below 50 hPa.

[0162] To create the dataset for developing the ML model, the WRF simulation was initialized at a certain moment and performed 9 simulations during a specific period. MSKF was called every 5 minutes throughout the simulation and output was generated at each call. Each simulation ran for 36 hours, of which the first 24 hours were used for training and the last 12 hours were used for validation. The total number of training samples was 114,444,000 and the offline validation set contained 57,024,000 samples.

[0163] Since there may be differences in offline performance, experiments were conducted to couple the WRF model with the machine learning model. The online effectiveness of this scheme was evaluated by comparing with the traditional WRF simulation results. The simulation was performed 4 times every 2 days, and each simulation lasted 168 hours.

[0164] Based on the above data set, the embodiment of the present invention is performed Figure 1 The steps shown are implemented. The Bi-LSTM model uses the Adam optimizer for network training and uses an initial learning rate of 0.003 to update the model parameters. In addition, the platform scheduler is implemented. When the loss does not decrease within 5 cycles, the learning rate is reduced by a factor of 0.5. The model is trained for 150 cycles with a batch size of 44,000.

[0165] The above is an example of one implementation method of this solution.

[0166] Corresponding to a meteorological convection simulation method based on machine learning provided in an embodiment of the present invention, see Fig.10 , the embodiment of the present invention also provides a structural block diagram of a meteorological convection simulation system based on machine learning, the meteorological convection simulation system includes: a first calculation unit 100, a second calculation unit 200, a judgment unit 300, a third calculation unit 400, and a first update unit 500;

[0167] The first calculation unit 100 is used to input grid parameters when the WRF model is integrated to time t to calculate the NCA.

[0168] The second calculation unit 200 is configured to calculate characteristic data of the convective cloud if the NCA is less than a time threshold.

[0169] In a specific implementation, the second calculation unit 200 is specifically used to: if the NCA is less than the time threshold, use a one-dimensional cloud model based on the multi-scale Kain-Fritsch algorithm, combined with temperature, dew point temperature, air pressure, wind speed and wind direction, to calculate the characteristic data of the convective cloud, and the characteristic data of the convective cloud includes at least: lifting condensation height, convective effective energy potential, average vertical velocity, cloud top height, cloud bottom height and entrainment rate.

[0170] The judging unit 300 is used to judge whether the characteristic data of the convective cloud meets the convection occurrence condition.

[0171] In a specific implementation, the judgment unit 300 is specifically used to: judge whether the convection effective energy bit exceeds the convection effective energy bit threshold; if the convection effective energy bit exceeds the convection effective energy bit threshold, determine that the convection occurrence condition is met; if the convection effective energy bit does not exceed the convection effective energy bit threshold, determine that the convection occurrence condition is not met.

[0172] The third calculation unit 400 is used to calculate the convection tendency and precipitation rate based on the specified machine learning model when the convection occurrence conditions are met, and update the average vertical velocity to execute the specified steps; wherein the convection tendency and precipitation rate are used to simulate meteorological convection in the WRF model, and the average vertical velocity is used to maintain the consistency of physical quantities.

[0173] In the specific implementation, the third computing unit 400 is specifically used for: when the convection occurrence conditions are met, at the grid points where the convection occurrence conditions have been met, using a multi-scale Kain-Fritsch algorithm based on a specified machine learning model to calculate the convection tendency and precipitation rate, wherein the specified machine learning model is a bidirectional long short-term memory model; updating the average vertical velocity and executing the specified steps.

[0174] The first updating unit 500 is used to update the average vertical velocity, set the convection tendency and precipitation rate to 0, and execute the specified steps when the convection occurrence condition is not met.

[0175] The specified steps include:

[0176] NCA decreases by 1 and t increases by 1, and the execution of the first calculation unit 100 is returned until the WRF model completes the iteration.

[0177] In an embodiment of the present invention, the WRF model and the machine learning model are coupled, and machine learning is applied to meteorological convection simulation to accurately predict the weather, thereby improving the effect of meteorological prediction.

[0178] Preferably, combined Fig.10 The meteorological convection simulation system further includes:

[0179] The second updating unit is used to update the average vertical velocity if NCA is greater than or equal to the time threshold, and the convective tendency and precipitation rate remain unchanged, and execute the above Fig.10 The specified steps are shown in .

[0180] In summary, the embodiments of the present invention provide a method and system for meteorological convection simulation based on machine learning, which couples the WRF model and the machine learning model, applies machine learning to meteorological convection simulation to accurately predict the weather, thereby improving the weather forecast effect.

[0181] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can refer to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system or system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiment. The system and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without creative work.

[0182] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0183] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A meteorological convection simulation method based on machine learning, characterized in that: The method comprises: The grid parameters are input into the Weather Research and Forecasting WRF model when it is integrated to time t to calculate the cloud relaxation time counter NCA; If the NCA is less than the time threshold, calculating characteristic data of the convective cloud; Determining whether the characteristic data of the convective cloud meets the convection occurrence condition; When the convection occurrence condition is met, the convection tendency and precipitation rate are calculated based on the specified machine learning model, and the average vertical velocity is updated, and the specified steps are executed; wherein the convection tendency and the precipitation rate are used to perform meteorological convection simulation in the WRF model, and the average vertical velocity is used to maintain the consistency of physical quantities; When the convection occurrence condition is not met, updating the average vertical speed, setting the convection tendency and the precipitation rate to 0, and executing the specified step; The designation step comprises: The NCA is reduced by 1 and t is increased by 1, and the step of inputting the grid parameters when the Weather Research and Forecasting WRF model is integrated to time t is returned to be executed until the WRF model completes the iteration.

2. The method according to claim 1, characterized in that If the NCA is less than the time threshold, characteristic data of the convective cloud are calculated, including: If the NCA is less than the time threshold, a one-dimensional cloud model is used based on the multi-scale Kain-Fritsch algorithm, and the characteristic data of the convective cloud are calculated in combination with the temperature, dew point temperature, air pressure, wind speed and wind direction. The characteristic data of the convective cloud include at least: lifting condensation height, convective effective energy potential, average vertical velocity, cloud top height, cloud base height and entrainment rate.

3. The method according to claim 2, characterized in that Determining whether the characteristic data of the convective cloud meets the convection occurrence condition includes: Determining whether the convection effective energy level exceeds a convection effective energy level threshold; If the convection effective energy level exceeds the convection effective energy level threshold, it is determined that the convection occurrence condition is met; If the convection effective energy level does not exceed the convection effective energy level threshold, it is determined that the convection occurrence condition is not met.

4. The method according to any one of claims 1 to 3, characterized in that: When the convection occurrence condition is met, the convection tendency and precipitation rate are calculated based on the specified machine learning model, and the average vertical velocity is updated, and the specified steps are executed, including: When the convection occurrence condition is met, at the grid point where the convection occurrence condition has been met, a multi-scale Kain-Fritsch algorithm based on a specified machine learning model is used to calculate the convection tendency and precipitation rate, wherein the specified machine learning model is a bidirectional long short-term memory model; Updates the average vertical velocity, taking the specified number of steps.

5. The method according to any one of claims 1 to 3, characterized in that: After calculating the cloud relaxation time counter NCA, it also includes: If the NCA is greater than or equal to the time threshold, the average vertical velocity is updated, the convective tendency and the precipitation rate remain unchanged, and the specified step is performed.

6. A meteorological convection simulation system based on machine learning, characterized in that: The system comprises: The first calculation unit is used to input the grid parameters when the Weather Research and Forecasting WRF model is integrated to time t to calculate the cloud relaxation time counter NCA; a second calculation unit, configured to calculate characteristic data of convective clouds if the NCA is less than a time threshold; A judging unit, used to judge whether the characteristic data of the convective cloud meets the convection occurrence condition; a third calculation unit, configured to calculate the convection tendency and precipitation rate based on a specified machine learning model and update the average vertical velocity to execute a specified step when the convection occurrence condition is met; wherein the convection tendency and precipitation rate are used to perform meteorological convection simulation in the WRF model, and the average vertical velocity is used to maintain the consistency of physical quantities; A first updating unit, configured to update the average vertical velocity, set the convection tendency and the precipitation rate to 0, and execute the specified step when the convection occurrence condition is not met; The designation step comprises: The NCA decreases by 1 and t increases by 1, and the first calculation unit is executed again until the WRF model completes the iteration.

7. The system according to claim 6, characterized in that The second calculation unit is specifically used to: if the NCA is less than a time threshold, use a one-dimensional cloud model based on a multi-scale Kain-Fritsch algorithm, combine temperature, dew point temperature, air pressure, wind speed and wind direction, and calculate the characteristic data of convective clouds, wherein the characteristic data of convective clouds include at least: lifting condensation height, convective effective energy potential, average vertical velocity, cloud top height, cloud bottom height and entrainment rate.

8. The system according to claim 7, characterized in that The judging unit is specifically used to: judge whether the convection effective energy level exceeds the convection effective energy level threshold; if the convection effective energy level exceeds the convection effective energy level threshold, determine that the convection occurrence condition is met; if the convection effective energy level does not exceed the convection effective energy level threshold, determine that the convection occurrence condition is not met.

9. The system according to any one of claims 6 to 8, characterized in that: The third calculation unit is specifically used for: when the convection occurrence condition is met, at the grid point where the convection occurrence condition has been met, using a multi-scale Kain-Fritsch algorithm based on a specified machine learning model to calculate the convection tendency and precipitation rate, wherein the specified machine learning model is a bidirectional long short-term memory model; updating the average vertical velocity and executing the specified steps.

10. The system according to any one of claims 6 to 8, characterized in that: Also includes: The second updating unit is used to update the average vertical speed if the NCA is greater than or equal to the time threshold, keep the convective tendency and the precipitation rate unchanged, and execute the specified step.