A parameterization method for high-order turbulence in numerical prediction based on machine learning

Abstract

The present invention discloses a numerical forecast high-order turbulence parameterization processing method based on machine learning, which uses large eddy simulation to extract the distribution characteristics of turbulent vertical motion in cloud physics process, and constructs a machine learning training data set containing atmospheric dynamics physical mechanism; according to the data set, the expectation maximization algorithm is used to obtain the double Gaussian probability density function of turbulent vertical motion, and the mixing weight of the obtained double Gaussian probability density function is fitted with the skewness of the vertical velocity in the large eddy simulation to obtain the probability density function mixing weight and skewness relationship required in the numerical forecast parameterization; according to the mixing weight and skewness relationship, the turbulent motion parameterization equation of the numerical forecast model is constructed. The present invention significantly improves the simulation accuracy of low cloud cloud amount related to the atmospheric physical process.

Description

Technical Field

[0001] The present invention relates to the technical field of atmospheric research and weather forecasting, and in particular to a method for parameterizing high-order turbulence in numerical forecasting based on machine learning. Background Art

[0002] Clouds have a vital impact on the Earth's climate system. The deviation in the current forecast precipitation in meteorological numerical values ​​is mainly related to the inaccurate numerical simulation of cloud physical processes. Numerical forecast models are limited by computing resources and cannot accurately analyze the turbulent motion that affects the formation and development of low clouds. They need to be replaced by parameterized processing methods, which makes it difficult to accurately reproduce the physical processes of low clouds in current numerical forecast models. Therefore, establishing an accurate parameterization method for boundary layer turbulence physical processes in numerical models has a vital impact on low cloud simulation.

[0003] In most models, the interaction between sub-grid turbulence, cloud amount, cloud microphysics, radiation and other processes needs to be constructed through multiple parameterization schemes, which leads to deviations in the simulation of low cloud amount and cloud water. In particular, for atmospheric turbulence processes, some models use the eddy diffusion mass flux method to unify the traditional dissipation form in the boundary layer with the mass flux method, thereby parameterizing the small-scale vertical motion in the boundary layer. This method shows a certain ability to simulate boundary layer clouds. In recent years, it has become a trend in the development of models to unify the parameterization of atmospheric boundary layer turbulence processes and cloud physics processes. Among them, the high-order turbulence closure method with a pre-assumed probability density function is used to deal with the small-scale processes in the boundary layer in the model, which has good internal consistency. Under this framework, the numerical model parameterization scheme unifies the description of the physical processes of boundary layer turbulence and cloud interaction through the pre-assumed probability density functions of sub-grid momentum and thermodynamic variables. This parameterization scheme has significantly improved the simulation of low clouds in global models, but studies have shown that the global low cloud amount is still underestimated compared with observations, indicating that the atmospheric turbulence parameterization scheme in the model still needs to be further improved.

[0004] The pre-assumed probability density function method used in turbulence parameterization is to pre-establish the calculation equation between the specified probability density function model and the statistical characteristics of turbulence in the parameterization scheme. Once the equation is established, the probability density function calculated at each time step and grid point can be used to close the higher-order terms in the turbulence prediction equation to calculate the cloud amount and cloud water content. In previous studies, various probability density function models suitable for exploring cloud characteristics have been discussed in depth and applied to global climate models. Including Gaussian distribution, trigonometric function distribution, β distribution, lognormal distribution, γ distribution and double Gaussian distribution. Existing research results show that the double Gaussian distribution can effectively describe the skewness of cloud parameters and can better represent the characteristics of shallow and deep convective clouds.

[0005] The double Gaussian probability density function based on three variables, vertical velocity, liquid water level temperature and total water mixing ratio, is currently used in the numerical model parameterization scheme, and has been proven to help improve low cloud simulation. However, the global low cloud amount simulated by the double Gaussian probability density function in the current numerical model is still underestimated compared with observations. This is because when a large vertical velocity skewed distribution appears in the boundary layer, the current double Gaussian function form still has a certain deviation in the skewed distribution of simulated boundary layer variables. In the numerical model parameterization scheme, it is necessary to pre-establish the hypothetical equation group of the double Gaussian function, so as to establish the relationship between the meteorological background field of each grid and the double Gaussian function model, and then obtain the small-scale turbulence distribution characteristics. Therefore, the pre-assumed double Gaussian probability density function has a particularly important impact on the simulation results. Studies have shown that the distribution of vertical velocity in the actual process is very complex, and the double Gaussian model established in the numerical model cannot well represent the skewed distribution structure, which is related to the high-order quantity of vertical velocity that represents the drastic change. Therefore, in order to accurately describe the turbulent motion in the numerical forecast model, it is necessary to try to reconstruct the probability density function equation group from the high-resolution large eddy simulation. Summary of the invention

[0006] The present invention aims to solve the technical problems existing in the related technologies of atmospheric turbulence parameterization at least to some extent.

[0007] One object of the present invention is to provide a method for parameterizing high-order turbulence for numerical forecasting based on machine learning. In view of the deficiency that the current numerical forecasting model cannot accurately describe the two vertical motions of turbulent rise and sink, machine learning is used to extract the vertical motion characteristics of turbulence from high-precision large eddy simulation, and to reconstruct the turbulence parameterization equations in numerical forecasting.

[0008] Another object of the present invention is to provide a numerical prediction high-order turbulence parameterization processing system based on machine learning.

[0009] In order to achieve the above-mentioned object, the present invention provides, on one hand, a method for numerical prediction of high-order turbulence parameterization based on machine learning, comprising:

[0010] Large eddy simulation was performed to simulate the physical processes of turbulent vertical motion of stratocumulus, shallow cumulus and cumulus clouds in the middle atmosphere using the large eddy model that takes into account the atmospheric wet process, and a high-resolution atmospheric vertical motion dataset was obtained.

[0011] Based on the high-resolution atmospheric vertical motion data set, the skewness data of the vertical velocity in the cloud physics process and the distribution characteristic data of the vertical velocity in the cloud physics process are calculated;

[0012] According to the distribution characteristic data of vertical velocity, the expectation maximization algorithm is used to calculate the mixing weight under the double Gaussian probability density distribution form, and the obtained mixing weight is fitted with the skewness data of vertical velocity to obtain the relationship between the mixing weight and skewness of the double Gaussian probability density function;

[0013] The parameterized equation of turbulent motion in the numerical prediction model is constructed based on the relationship between the mixing weight and skewness of the double Gaussian probability density function.

[0014] A further preferred technical solution of the present invention is that, according to the high-resolution atmospheric vertical motion data set, the skewness data of the vertical velocity in the cloud physics process is calculated. The specific method is to use the average vertical velocity of the entire layer and the instantaneous vertical velocity variables of each grid in the high-resolution atmospheric vertical motion data to calculate the turbulent average disturbance variable, and then calculate the skewness data of the vertical velocity through the turbulent average disturbance variable;

[0015] At the same time, based on the high-resolution atmospheric vertical motion dataset, the distribution characteristic data of the vertical velocity in the cloud physics process are calculated. The specific method is to use the kernel density distribution method to calculate the probability density distribution function of the vertical velocity in the atmospheric vertical motion dataset as the distribution characteristic data of the vertical velocity.

[0016] Preferably, the large eddy simulation is performed by using a Python architecture to consider a large eddy model of atmospheric wet processes, using inelastic motion equations in the large eddy model, and using a wet isentropic surface to calculate the change in buoyancy acceleration caused by atmospheric density fluctuations, which is used to calculate atmospheric turbulent motion characteristic samples of different types of cloud physical processes. Each sample is selected by taking each altitude layer at each moment in a high-resolution atmospheric vertical motion dataset as an atmospheric turbulent motion characteristic sample for training an expectation-maximization algorithm.

[0017] Preferably, according to the distribution characteristic data of the vertical velocity, the expectation maximization algorithm is used to calculate the mixing weight in the form of double Gaussian probability density distribution, and the obtained mixing weight is fitted with the skewness data of the vertical velocity to obtain the relationship between the mixing weight and skewness of the double Gaussian probability density function; the specific method is:

[0018] For the atmospheric turbulence motion characteristic samples, the expectation maximization algorithm is used to calculate the double Gaussian probability density function that best matches the distribution characteristic data of the vertical velocity, and a mixing weight in the form of a double Gaussian probability density distribution is obtained;

[0019] The mixing weights in the form of double Gaussian probability density distribution obtained at each altitude layer at each moment are fitted with the skewness data of the vertical velocity at the corresponding altitude layer at the corresponding moment to obtain the relationship between the mixing weights and skewness of the double Gaussian probability density function.

[0020] Preferably, when using the expectation-maximization algorithm to obtain the mixing weights in the form of a double Gaussian probability density distribution, the ascending vertical velocity in the atmospheric turbulent motion characteristic sample is used as the first Gaussian function, and the sinking vertical velocity is used as the second Gaussian function. The expectation-maximization algorithm is used to obtain the mixing weights in the form of a double Gaussian probability density distribution that is most similar to the distribution characteristic data of the vertical velocity through 10,000 iterations.

[0021] Preferably, the parameterized equation of turbulence motion in the numerical prediction model is constructed according to the relationship between the mixed weight and skewness of the double Gaussian probability density function, specifically:

[0022] Construct relations for the five variables of the double Gaussian probability density function to be solved in the numerical forecast model;

[0023] The five variables are double Gaussian mixture weight variables, the first Gaussian mean variable and the second Gaussian mean variable, as well as the first Gaussian standard deviation variable and the second Gaussian standard deviation variable; the relationship between the variables is:

[0024]

[0025]

[0026]

[0027]

[0028]

[0029] in, is the double Gaussian mixture weight, and are the first Gaussian mean variable and the second Gaussian mean variable, respectively. and is the standard deviation variable of the first Gaussian and the standard deviation variable of the second Gaussian, is the skewness of the vertical velocity.

[0030] Another aspect of the present invention provides a numerical prediction high-order turbulence parameterization processing system based on machine learning, comprising:

[0031] The dataset construction module is used to perform large eddy simulation. It uses the large eddy model that takes into account the atmospheric wet process to simulate the physical processes of turbulent vertical motion in the middle atmosphere of stratocumulus, shallow cumulus and cumulus clouds, respectively, to obtain a high-resolution atmospheric vertical motion dataset.

[0032] The turbulent vertical motion probability density function acquisition module is used to calculate the skewness data of the vertical velocity in the cloud physics process and the distribution characteristic data of the vertical velocity in the cloud physics process based on the high-resolution atmospheric vertical motion data set;

[0033] It is used to calculate the mixing weight under the double Gaussian probability density distribution form according to the distribution characteristic data of the vertical velocity by using the expectation maximization algorithm, and the obtained mixing weight is fitted with the skewness data of the vertical velocity to obtain the relationship between the mixing weight and skewness of the double Gaussian probability density function;

[0034] The turbulence motion parameterized equation acquisition module is used to construct the turbulence motion parameterized equation of the numerical prediction model according to the relationship between the mixing weight and skewness of the double Gaussian probability density function.

[0035] Another aspect of the present invention provides a non-transitory computer-readable storage medium having computer instructions stored thereon, which enable a computer to execute the above-mentioned numerical prediction high-order turbulence parameterization processing method based on machine learning.

[0036] Yet another aspect of the present invention provides an electronic device, comprising: a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus, and the processor calls the logic instructions in the memory to execute the above-mentioned numerical prediction high-order turbulence parameterization processing method based on machine learning.

[0037] On the other hand, the present invention provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer executes the above-mentioned numerical prediction high-order turbulence parameterization processing method based on machine learning.

[0038] Beneficial effects: Compared with the original turbulence parameterization scheme, the present invention has the following advantages:

[0039] 1. In order to solve the problem of underestimation of high-order disturbance terms of turbulence in the original numerical prediction method, the technical solution of the present invention extracts turbulence distribution characteristics from large eddy simulation that can accurately describe turbulent motion, and provides a data set for the numerical prediction parameterization scheme.

[0040] 2. Use machine learning algorithms to extract accurate turbulence distribution functions from large eddy simulation data sets and establish more accurate turbulence motion parameterization equations than the original scheme.

[0041] 3. The parameterization scheme based on machine learning is coupled into the numerical forecast model, which significantly improves the low cloud amount in the numerical forecast model.

[0042] In summary, in order to make up for the deficiency of the current numerical forecast model in accurately describing the two vertical motions of turbulent rise and sink, the present invention uses machine learning to extract the vertical motion characteristics of turbulence from high-precision large eddy simulation and reconstructs the turbulence parameterization equation in numerical forecast. It improves the accuracy of numerical model simulation of small-scale turbulent motion and the simulation effect of numerical model on low clouds, thereby improving the accuracy of numerical forecast. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a flow chart of the method for numerical prediction of high-order turbulence parameterization based on machine learning according to the present invention;

[0044] Figure 2 It is a comparison diagram of the projection profile diagnosis results of the joint probability density distribution based on the 7-8 hour average height of the RICO cumulus cloud case at an altitude of z=450m and 650m in the embodiment. DETAILED DESCRIPTION

[0045] 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 in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments, and they should not be understood as limitations on the present invention. 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. In the description of the present invention, it should be understood that the terms used are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0046] Combine the following Figure 1-Figure 2 The present invention describes a numerical prediction high-order turbulence parameterization processing method based on machine learning, a numerical prediction high-order turbulence parameterization processing system based on machine learning, a medium, an electronic device and a computer program.

[0047] Embodiment 1: This embodiment provides a method for numerical prediction of high-order turbulence parameterization based on machine learning.

[0048] In reality, the distribution of vertical velocity is very complex, and the double Gaussian model established in the numerical model cannot well represent the skewed distribution structure. Studies have shown that this structure is related to the background of high-order quantities of vertical velocity that represent drastic changes.

[0049] In recent years, with the rapid development of artificial intelligence, the use of machine learning to model atmospheric turbulence has gradually become a hot topic of research. Among them, the expectation maximization algorithm can extract the probability density function distribution of variables from large eddy simulation data, and shows good prospects in improving the accuracy of modeling given functional forms. The expectation maximization algorithm is a machine learning algorithm that can find the maximum likelihood estimate of a given functional form. The algorithm can generate a double Gaussian distribution form that is most similar to the turbulence distribution of large eddy simulation. This method has good application value for constructing a probability density function equation group describing turbulent motion in numerical forecasting models.

[0050] In order to make up for the deficiency of the current numerical forecast model in accurately describing the two vertical motions of turbulent rise and sink, this embodiment uses machine learning to extract the vertical motion characteristics of turbulence from high-precision large eddy simulations and reconstruct the turbulence parameterization equations in numerical forecasts. This improves the accuracy of numerical models in simulating small-scale turbulent motion and the simulation effect of numerical models on low clouds, thereby improving the accuracy of numerical forecasts.

[0051] This embodiment mainly includes two parts: one is to extract the double Gaussian function characteristics of turbulent motion in large eddy simulation based on machine learning, and the other is to construct the double Gaussian function solution equation in numerical forecasting based on the double Gaussian function characteristics of machine learning. The specific steps are as follows:

[0052] S1. Perform large eddy simulation, using the large eddy model that considers the atmospheric wet process to simulate the physical processes of turbulent vertical motion in the middle atmosphere of stratocumulus, shallow cumulus and cumulus, and obtain a high-resolution atmospheric vertical motion dataset;

[0053] In this example, the Python framework large eddy simulation model (PyCLES) is used to establish a training data set for turbulent vertical motion characteristics. This model framework is based on the formula of the elastic equation and can accurately simulate deep convection, which has certain advantages for clouds with frequent phase changes. Specifically, in this example, three different cloud type observation initial field settings are selected for large eddy simulation, namely: stratocumulus observation (DYCOMS-II RF01), shallow cumulus observation (BOMEX), and cumulus observation (RICO), and the specific settings are as follows:

[0054] 1. DYCOMS-II RF01 This flight observation experiment was conducted at 122°W, 31°N on the coast of Southern California from 06:00 to 15:00 on July 10, 2001. Its design purpose is to measure the marine stratocumulus cloud area to understand the stratocumulus cloud structure. In this embodiment, PycLES is used to simulate DYCOMS RF01 with a vertical resolution of 5m, a horizontal resolution of 35m, and a time resolution of 1s. The simulation area size is 3396m×3396m, and the vertical height is 1500m. The cloud droplet number concentration is set to 120cm-3 The simulation lasted 4 hours, and the average result of the last hour was selected.

[0055] 2. The shallow cumulus example is the BOMEX experiment project, which was conducted from May 1 to July 28, 1969. The main goal of the project was to measure the exchange rate of heat, moisture and momentum between the tropical ocean and the atmosphere in a 500 square kilometer area under the trade wind environment. The simulation settings are 40m vertical resolution, 100m horizontal resolution, and 1s time resolution. The simulation area size is 6.4km×6.4km, and the vertical height is 3km. The simulation lasts for 6 hours, and the average result of the last hour is selected.

[0056] 3. RICO is a marine cumulonimbus rain observation project conducted over the western Atlantic Ocean from November 2004 to January 2005. Its purpose is to measure the processes related to the formation of rain in shallow cumulonimbus clouds and how precipitation subsequently changes the structure of trade wind clouds. The statistical and research settings are the same, with a vertical resolution of 40m, a horizontal resolution of 100m, and a time resolution of 1s. The simulation area size is 12.8km×12.8km, and the vertical height is 6km. The cloud droplet number concentration is set to 70cm -3 The simulation lasts for 24 hours, and the average result of the last hour is selected.

[0057] S2. Calculate the skewness data of the vertical velocity in the cloud physics process and the distribution characteristic data of the vertical velocity in the cloud physics process based on the high-resolution atmospheric vertical motion data set;

[0058] According to the distribution characteristic data of vertical velocity, the expectation maximization algorithm is used to calculate the mixing weight under the double Gaussian probability density distribution form. The obtained mixing weight is fitted with the skewness data of vertical velocity to obtain the relationship between the mixing weight and skewness of the double Gaussian probability density function:

[0059] The two Gaussian distributions on the left and right in the double Gaussian probability density distribution represent downdrafts and updrafts, respectively, which can be regarded as a cluster analysis problem in machine learning. This requires separating the distribution of updrafts and downdrafts from the entire large eddy simulation area.

[0060] Using the turbulent motion data obtained by large eddy simulation, the mixed weights in the form of double Gaussian probability density distribution are obtained through the expectation maximization algorithm in the machine learning algorithm. The main process is as follows: Figure 1 As shown, each altitude layer at each moment in the LES simulation results is sampled, and the parameters of the double Gaussian probability density function that best matches the LES data are calculated using a maximization algorithm to obtain a mixing weight in the form of a double Gaussian probability density distribution, which is used as a training sample;

[0061] The mixed weight of the double Gaussian probability density distribution obtained at each altitude layer at each moment is fitted with the skewness of the vertical velocity at the corresponding altitude layer at the corresponding moment to obtain the relationship between the probability density function weight and the skewness. The probability density function parameterization equation in the numerical forecast model parameterization scheme is improved according to the obtained relationship.

[0062] In this embodiment, a machine learning algorithm is used to iterate 10,000 times the vertical velocity probability density function calculated from the large eddy simulation data of each layer at each time to obtain the most accurate double Gaussian mixing weight compared with the turbulence obtained by the large eddy simulation. Each height at each time in the large eddy simulation is taken as a sample, and a total of 3,300 samples are obtained after training the three typical cloud case data.

[0063] S3. According to the relationship between the mixed weight and skewness of the double Gaussian probability density function, the parameterized equation of turbulent motion in the numerical prediction model is constructed:

[0064] In the numerical prediction model, the five variables of the double Gaussian probability density function to be solved are the double Gaussian mixture weight variable, the first Gaussian mean variable and the second Gaussian mean variable, and the standard deviation variable of the first Gaussian and the standard deviation variable of the second Gaussian;

[0065] Therefore, the new double Gaussian probability density function parameterization formula established in this embodiment is:

[0066]

[0067]

[0068]

[0069]

[0070]

[0071] in, is the double Gaussian mixture weight variable, and are the first Gaussian mean variable and the second Gaussian mean variable, respectively. and is the standard deviation variable of the first Gaussian and the standard deviation variable of the second Gaussian, is the skewness of the vertical velocity.

[0072] This embodiment significantly improves the underestimation problem of high-order turbulence in the original numerical prediction parameterization scheme. Figure 2The cross-sectional diagram of the double Gaussian probability density function distribution of thermodynamic variables in the cloud at different altitudes for the RICO case. The results show that when the average altitude is z=450m and z=650m, the sub-grid turbulence distribution in the numerical model is closer to the high-precision large eddy simulation after the new scheme of this embodiment is implemented. The original scheme underestimates the supersaturated area in the projection of the joint probability density function of the total water mixing ratio and the liquid water level temperature in the thermodynamic variables, resulting in an underestimation of the simulated low cloud amount.

[0073] The improved parameterization scheme of this embodiment further improves the simulation accuracy of global low cloud amount, especially the simulation accuracy of low cloud area on the west coast of the continent is significantly improved. By comparing the global distribution of the 5-year average low cloud amount simulated by the global model, the results show that compared with satellite observations, the original scheme significantly underestimates the low cloud amount in the transition zone from stratocumulus to cumulus on the coast. Compared with the original scheme, the new scheme of this embodiment increases the cloud amount in the transition zone from stratocumulus to cumulus by about 30%-40%, which is closer to the satellite observation results. This result is of great significance to the global numerical forecast model.

[0074] Embodiment 2: This embodiment provides a numerical prediction high-order turbulence parameterization processing system based on machine learning, comprising:

[0075] A data set construction module is used to perform large eddy simulation and extract turbulent vertical motion data in large eddy simulation to construct a data set;

[0076] The turbulent vertical motion probability density function acquisition module is used to obtain the mixing weight of the double Gaussian probability density distribution form of turbulent vertical motion using the expectation maximization algorithm according to the data set, and fit the mixing weight of the double Gaussian probability density distribution form with the skewness of the vertical velocity in the large eddy simulation to obtain the relationship between the probability density function weight and the skewness;

[0077] The module for acquiring the parameterized equation of turbulence motion is used to construct the parameterized equation of turbulence motion in the numerical prediction model according to the relationship between the probability density function weight and the skewness.

[0078] Embodiment 3: This embodiment provides a non-transitory computer-readable storage medium having computer instructions stored thereon, the computer instructions causing a computer to execute a numerical prediction high-order turbulence parameterization processing method based on machine learning, the method comprising the following steps:

[0079] S1. Perform large eddy simulation and extract turbulent vertical motion data in large eddy simulation to construct a data set;

[0080] S2. According to the data set, the expectation maximization algorithm is used to obtain the mixing weight of the double Gaussian probability density distribution form of the turbulent vertical motion, and the mixing weight of the double Gaussian probability density distribution form is fitted with the skewness of the vertical velocity in the large eddy simulation to obtain the relationship between the probability density function weight and the skewness;

[0081] S3. Construct the parameterized equation of turbulent motion in the numerical prediction model based on the relationship between the probability density function weight and the skewness.

[0082] Embodiment 4: This embodiment provides an electronic device, which may include: a processor, a communications interface, a memory, and a communication bus, wherein the processor, the communications interface, and the memory communicate with each other via the communication bus. The processor may call logic instructions in the memory to execute a numerical prediction high-order turbulence parameterization processing method based on machine learning, the method comprising the following steps:

[0083] S1. Perform large eddy simulation and extract turbulent vertical motion data in large eddy simulation to construct a data set;

[0084] S2. According to the data set, the expectation maximization algorithm is used to obtain the mixing weight of the double Gaussian probability density distribution form of the turbulent vertical motion, and the mixing weight of the double Gaussian probability density distribution form is fitted with the skewness of the vertical velocity in the large eddy simulation to obtain the relationship between the probability density function weight and the skewness;

[0085] S3. Construct the parameterized equation of turbulent motion in the numerical prediction model based on the relationship between the probability density function weight and the skewness.

[0086] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program code.

[0087] Embodiment 5: This embodiment provides a computer program product, which includes a computer program. The computer program can be stored in a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute a numerical prediction high-order turbulence parameterization processing method based on machine learning, which includes the following steps:

[0088] S1. Perform large eddy simulation and extract turbulent vertical motion data in large eddy simulation to construct a data set;

[0089] S2. According to the data set, the expectation maximization algorithm is used to obtain the mixing weight of the double Gaussian probability density distribution form of the turbulent vertical motion, and the mixing weight of the double Gaussian probability density distribution form is fitted with the skewness of the vertical velocity in the large eddy simulation to obtain the relationship between the probability density function weight and the skewness;

[0090] S3. Construct the parameterized equation of turbulent motion in the numerical prediction model based on the relationship between the probability density function weight and the skewness.

[0091] The device embodiments described above are merely illustrative, 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 paying creative labor.

[0092] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0093] 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 method for parameterizing high-order turbulence for numerical prediction based on machine learning, characterized in that: include: Large eddy simulation was performed to simulate the physical processes of turbulent vertical motion of stratocumulus, shallow cumulus and cumulus clouds in the middle atmosphere using the large eddy model that takes into account the atmospheric wet process, and a high-resolution atmospheric vertical motion dataset was obtained. Based on the high-resolution atmospheric vertical motion data set, the skewness data of the vertical velocity in the cloud physics process and the distribution characteristic data of the vertical velocity in the cloud physics process are calculated; According to the distribution characteristic data of vertical velocity, the expectation maximization algorithm is used to calculate the mixing weight under the double Gaussian probability density distribution form, and the obtained mixing weight is fitted with the skewness data of vertical velocity to obtain the relationship between the mixing weight and skewness of the double Gaussian probability density function; The parameterized equation of turbulent motion in the numerical prediction model is constructed based on the relationship between the mixed weight and skewness of the double Gaussian probability density function; specifically: Construct relations for the five variables of the double Gaussian probability density function to be solved in the numerical forecast model; The five variables are double Gaussian mixture weight variables, the first Gaussian mean variable and the second Gaussian mean variable, as well as the first Gaussian standard deviation variable and the second Gaussian standard deviation variable; the relationship between the variables is: Among them, a is the weight of the double Gaussian mixture, and are the first Gaussian mean variable and the second Gaussian mean variable, respectively. and Sk is the standard deviation variable of the first Gaussian and the standard deviation variable of the second Gaussian, w is the skewness of the vertical velocity.

2. The method for numerical prediction of high-order turbulence parameterization based on machine learning according to claim 1, characterized in that: The method of calculating the skewness data of the vertical velocity in the cloud physics process according to the high-resolution atmospheric vertical motion data set is as follows: using the average vertical velocity of the entire layer and the instantaneous vertical velocity variables of each grid in the high-resolution atmospheric vertical motion data, calculating the turbulence average disturbance variable, and then calculating the skewness data of the vertical velocity through the turbulence average disturbance variable; At the same time, based on the high-resolution atmospheric vertical motion dataset, the distribution characteristic data of the vertical velocity in the cloud physics process are calculated. The specific method is to use the kernel density distribution method to calculate the probability density distribution function of the vertical velocity in the atmospheric vertical motion dataset as the distribution characteristic data of the vertical velocity.

3. The method for numerical prediction of high-order turbulence parameterization based on machine learning according to claim 1, characterized in that: The large eddy simulation is performed by using a Python architecture to consider a large eddy model of atmospheric wet processes, using an inelastic motion equation in the large eddy model, and using a wet isentropic surface to calculate the change in buoyancy acceleration caused by atmospheric density fluctuations, which is used to calculate atmospheric turbulent motion characteristic samples of different types of cloud physical processes. Each sample is selected by taking each altitude layer at each moment in a high-resolution atmospheric vertical motion dataset as an atmospheric turbulent motion characteristic sample for training an expectation-maximization algorithm.

4. The method for numerical prediction of high-order turbulence parameterization based on machine learning according to claim 3 is characterized in that: According to the distribution characteristic data of vertical velocity, the expectation maximization algorithm is used to calculate the mixing weight under the double Gaussian probability density distribution form, and the obtained mixing weight is fitted with the skewness data of vertical velocity to obtain the relationship between the mixing weight and skewness of the double Gaussian probability density function; the specific method is: For the atmospheric turbulence motion characteristic samples, the expectation maximization algorithm is used to calculate the double Gaussian probability density function that best matches the distribution characteristic data of the vertical velocity, and a mixing weight in the form of a double Gaussian probability density distribution is obtained; The mixing weights in the form of double Gaussian probability density distribution obtained at each altitude layer at each moment are fitted with the skewness data of the vertical velocity at the corresponding altitude layer at the corresponding moment to obtain the relationship between the mixing weights and skewness of the double Gaussian probability density function.

5. The method for numerical prediction of high-order turbulence parameterization based on machine learning according to claim 4, characterized in that: When the expectation maximization algorithm is used to obtain the mixing weights in the form of double Gaussian probability density distribution, the ascending vertical velocity in the atmospheric turbulent motion characteristic samples is used as the first Gaussian function, and the sinking vertical velocity is used as the second Gaussian function. The expectation maximization algorithm is used through 10,000 iterations to obtain the mixing weights in the form of double Gaussian probability density distribution that is most similar to the distribution characteristic data of the vertical velocity.

6. A numerical prediction high-order turbulence parameterization processing system based on machine learning, characterized in that: include: The dataset construction module is used to perform large eddy simulation. It uses the large eddy model that takes into account the atmospheric wet process to simulate the physical processes of turbulent vertical motion in the middle atmosphere of stratocumulus, shallow cumulus and cumulus clouds, respectively, to obtain a high-resolution atmospheric vertical motion dataset. The turbulent vertical motion probability density function acquisition module is used to calculate the skewness data of the vertical velocity in the cloud physics process and the distribution characteristic data of the vertical velocity in the cloud physics process based on the high-resolution atmospheric vertical motion data set; It is used to calculate the mixing weight under the double Gaussian probability density distribution form according to the distribution characteristic data of the vertical velocity by using the expectation maximization algorithm, and the obtained mixing weight is fitted with the skewness data of the vertical velocity to obtain the relationship between the mixing weight and skewness of the double Gaussian probability density function; The turbulence motion parameterized equation acquisition module is used to construct the turbulence motion parameterized equation of the numerical prediction model according to the mixed weight and skewness relationship of the double Gaussian probability density function, specifically: Construct relations for the five variables of the double Gaussian probability density function to be solved in the numerical forecast model; The five variables are double Gaussian mixture weight variables, the first Gaussian mean variable and the second Gaussian mean variable, as well as the first Gaussian standard deviation variable and the second Gaussian standard deviation variable; the relationship between the variables is: Among them, α is the weight of the double Gaussian mixture, and are the first Gaussian mean variable and the second Gaussian mean variable, respectively. and Sk is the standard deviation variable of the first Gaussian and the standard deviation variable of the second Gaussian, w is the skewness of the vertical velocity.

7. A non-transitory computer-readable storage medium, characterized in that: Computer instructions are stored thereon, and the computer instructions enable the computer to execute the numerical prediction high-order turbulence parameterization processing method based on machine learning as described in any one of claims 1 to 5.

8. An electronic device, characterized in that: include: A processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus, and the processor calls the logic instructions in the memory to execute the numerical prediction high-order turbulence parameterization processing method based on machine learning as described in any one of claims 1 to 5.

9. A computer program product, characterized in that The computer program product includes a computer program stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer executes the numerical prediction high-order turbulence parameterization processing method based on machine learning as described in any one of claims 1 to 5.

Images