Microphysical quantity prediction method and model training method after cloud entrainment, mixing and evaporation process

Abstract

The present invention relates to the field of meteorological forecasting technology, and in particular to a method for predicting microphysical quantities after a mixed evaporation process in a cloud, a method for training a machine learning model, and a computer program product. The machine learning model training method obtains a parameterized data set of a mixed evaporation process in a cloud based on an explicit mixed bubble model, uses the microphysical quantities in the cloud and the environmental quantities outside the cloud of the parameterized data set as characteristic quantities, and uses the number concentration and discreteness as prediction quantities to train the machine learning model, thereby realizing the prediction of microphysical quantities after a mixed evaporation process in a cloud. The microphysical quantity prediction method does not need to establish a specific physical relationship between the characteristic quantity and the prediction quantity, and can obtain more accurate microphysical quantities after a mixed evaporation process in a cloud compared to traditional parameterization schemes.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological forecasting, and in particular to a method for predicting microphysical quantities after a cloud entrainment, mixing and evaporation process, a training method for a machine learning model for predicting microphysical quantities after a cloud entrainment, mixing and evaporation process, and a computer program product. Background Art

[0002] Clouds cover about 2 / 3 of the Earth's surface and are an important part of the Earth-atmosphere system. Accurate simulation of cloud precipitation is a bottleneck problem in meteorological model research. At present, there are great uncertainties and differences in the simulation of cloud feedback in different meteorological models, such as weather models and climate models, which seriously affect the accuracy of meteorological forecasts. Aerosol-cloud is the most uncertain part of the various factors affecting meteorological changes. In the study of aerosol-cloud interactions, the entrainment mixing evaporation process is one of the most uncertain physical processes. Therefore, it is very important to study and accurately identify the entrainment mixing evaporation process in clouds.

[0003] Studying the entrainment, mixing and evaporation process in clouds requires a deep understanding of the microphysics of clouds. The characteristics of cloud microphysics are usually described by microphysical quantities such as the phase state, scale, concentration, droplet spectrum and water content of cloud particles. Existing technologies use cloud microphysics schemes, that is, a parametric method to describe cloud microphysical processes, to study the physical properties of clouds and the interaction between clouds and the atmospheric environment. The entrainment and mixing mechanisms in clouds under different environmental conditions are different, but in the microphysics schemes of weather / climate models, for the convenience of calculation, it can only be simply assumed to be some extreme entrainment and mixing mechanisms, such as vertical mixing mechanisms, overall mixing mechanisms and uniform / inhomogeneous mixing mechanisms, among which uniform / inhomogeneous mixing mechanisms are the most widely used.

[0004] The uniform / non-uniform mixing mechanism is based on the different effects of the entrained air on the cloud droplet spectrum, and the entrainment mixing process is divided into a uniform mixing process and an extremely non-uniform mixing process. In the uniform entrainment mixing process, when the ambient air is entrained into the cloud, it is quickly and uniformly mixed with all the cloud droplets. All the cloud droplets are in the same unsaturated state and evaporate at the same time. At this time, the cloud droplet size is reduced, but the cloud droplet number concentration remains unchanged, and the volume average radius is positively correlated with the water content, and the cloud droplet spectrum moves toward the small droplet direction. In the extremely non-uniform entrainment mixing process, when the ambient air is entrained into the cloud, only a part of the cloud droplets close to the entrained air will completely evaporate, while the part of the cloud droplets far away from the entrained air will not be affected. At this time, the cloud droplet size remains unchanged, the cloud droplet number concentration decreases, and the volume average radius remains unchanged with the water content, and the cloud droplet spectrum shifts downward.

[0005] The prior art “Exploring parameterization for turbulent entrainment-mixing processes in clouds” discloses quantitatively characterizing the entrainment mixing mechanism from a microphysical perspective using uniform mixing percentage (ψ).

[0006]

[0007] Among them, N c is the cloud droplet number concentration after entrainment, mixing and evaporation, N h and N i are the cloud droplet number concentrations after uniform entrainment and mixing evaporation and extreme non-uniform entrainment and mixing evaporation, respectively. The calculation formula is:

[0008]

[0009] N h =N a χ (3)

[0010] Among them, r vc is the volume average radius after the entrainment mixing evaporation process, r va is the volume average radius after the extremely non-uniform entrainment mixing and evaporation process. N a is the adiabatic number concentration, and χ is the proportion of adiabatic cloud after entrainment mixing occurs. The existing technology defines the range of ψ as 0 to 1, where 0 represents an extremely non-uniform entrainment mixing mechanism and 1 represents a uniform entrainment mixing mechanism. This quantity can not only characterize the two extreme states, but also describe the intermediate state. The larger the ψ value, the more uniform the entrainment mixing process.

[0011] In addition, the prior art "Examination of turbulent entrainment-mixing mechanisms using a combined approach" discloses the transition scale number (N L ), characterize the entrainment mixing mechanism from the dynamic aspect:

[0012]

[0013] Where L * and η is the transition length and the Kolmogorov length scale:

[0014] L*=ε 1*2 τ r 3 / 2 , (5)

[0015]

[0016] In equations (5) and (6), ε is the turbulent dissipation rate, τr is the microphysical response time of cloud droplets, and ν is the kinematic viscosity. L The larger it is, the more uniform the entrainment and mixing process is.

[0017] The prior art "Parameterizations of entrainment-mixing mechanisms and their effects on cloud droplet spectral width based on numerical simulations" discloses the uniform mixing percentage ψ and the transition scale number N L Related relationships:

[0018]

[0019] Cloud droplet number concentration N in response to entrainment, mixing and evaporation processes in weather / climate models c It is expressed as:

[0020]

[0021] where q c0 and q c They represent the water vapor mixing ratio before and after the entrainment mixing and evaporation process. Generally speaking, the weather model is often set to uniform entrainment mixing, ψ = 1; the climate model is often set to extreme non-uniform entrainment mixing, ψ = 0.

[0022] Based on the above physical formulas (1) to (8), the uniform mixing percentage at the current moment can be diagnosed in real time according to the cloud droplet properties and environmental conditions at each moment, thereby more accurately characterizing the number concentration of cloud droplets.

[0023] In addition, since the cloud droplet spectrum does not change in the case of uniform entrainment mixing and extreme non-uniform entrainment mixing, the cloud droplet spectrum discreteness (dis) in the intermediate state is expressed as:

[0024] dis=0.03+0.008ψ-7.17×10 -5 ψ 2 , (9)

[0025] The prior art calculates the transition scale number N according to formulas (4) to (6): L , and then according to formula (7) from N L The uniform mixing percentage ψ is calculated, and finally the cloud droplet number concentration N after the entrainment mixing process is calculated from ψ c and cloud droplet spectrum dispersion dis. The prior art obtains the cloud droplet number concentration N based on the above traditional parameterization scheme. cThe parameterization results of these two microphysical quantities, i.e., cloud droplet spectrum discreteness dis, are used to improve the accuracy of cloud precipitation process prediction.

[0026] In the existing research on uniform / inhomogeneous mixing mechanism, there is a great controversy about which is the dominant one, the uniform entrainment mixing process or the extremely inhomogeneous entrainment mixing process. In addition, since the existing technology relies on many assumptions to describe the entrainment mixing process, the physical formula used to describe it is not accurate enough, which makes it impossible to accurately determine the cloud droplet number concentration N after the entrainment mixing evaporation process. c and the value of cloud droplet spectrum discreteness dis, which affects the accuracy of meteorological forecasts. Summary of the invention

[0027] The object of the present invention is to provide a method for more accurately predicting microphysical quantities after the entrainment, mixing and evaporation process in clouds.

[0028] In order to achieve the above invention objectives, in a first aspect, the present invention provides a method for predicting microphysical quantities after a cloud entrainment mixing and evaporation process, comprising the following steps:

[0029] A parameterized data set of cloud entrainment mixing and evaporation process is obtained based on an explicit mixing bubble model, wherein the parameterized data set includes microphysical quantities inside the cloud and environmental quantities outside the cloud;

[0030] The parameterized data set is input into a pre-trained machine learning model, and the corresponding number concentration Nc and discreteness dis output by the machine learning model are used as the predicted microphysical quantities after the entrainment mixing and evaporation process in the cloud.

[0031] Preferably, the cloud microphysical quantities include: adiabatic water content LWCa, adiabatic number concentration Na, adiabatic volume mean radius r va , and multiple water contents LWCc after entrainment, mixing and evaporation.

[0032] Preferably, the extra-cloud environmental quantities include: multiple ones of: turbulence dissipation rate ε, ambient air temperature Te, ambient air relative humidity RHe, and ambient air water vapor mixing ratio qve.

[0033] Preferably, the parameterized data set for obtaining the entrainment mixing evaporation process in the cloud based on the explicit mixing bubble model is specifically to obtain the model simulation area and its initial parameter settings, the initial parameters including the model scale Dm, vertical velocity w, adiabatic number concentration Na, the ratio of entrained ambient air L fraction , ambient air relative humidity RHe and turbulent dissipation rate ε; multiple groups of sensitivity tests are implemented based on the initial parameters to obtain the parameter changes of the cloud entrainment mixing evaporation process, and the parameters before and after the cloud entrainment mixing evaporation process are recorded in the parameterized data set respectively.

[0034] In a second aspect, the present invention provides a machine learning model training method for predicting microphysical quantities after entrainment, mixing and evaporation in clouds, comprising the following steps:

[0035] Obtain a parameterized data set, use the cloud microphysical quantities and cloud environment quantities processed as any one of claims 1 to 5 in the parameterized data set as characteristic quantities, and use the number concentration Nc and the dispersion dis after the entrainment mixing evaporation process as prediction quantities; divide all characteristic quantities and prediction quantities into two groups, which are respectively used as a training set and a validation set;

[0036] Train the machine learning model, use the feature quantity of the training set as the model input, and the prediction quantity as the model output to train the model, so as to obtain the ability to predict the microphysical quantities after the entrainment mixing evaporation process in the cloud through the input microphysical quantities in the cloud and the environmental quantities outside the cloud;

[0037] Model evaluation: Use the validation set to evaluate the microphysical quantity prediction performance of the trained machine learning model, that is, compare the microphysical quantities predicted by the machine learning model with the parameterized results of the microphysical quantities.

[0038] Preferably, the machine learning model includes: extreme learning machine ELM, multi-label learning MLL, support vector machine SVM, decision tree DT, random forest RF, gradient boosting tree GBDT or lightweight gradient boosting machine LightGBM.

[0039] Preferably, an additive explanatory model based on SHAP is established. For each feature quantity of the parameterized data set, the additive explanatory model generates a prediction value for each prediction sample based on all feature quantities, and the SHAP value of each feature quantity is used to characterize the contribution of the feature quantity to the prediction quantity. The larger the absolute value of the SHAP value, the greater the contribution to the prediction value.

[0040] Preferably, the model evaluation is specifically as follows: taking the microphysical quantities calculated based on the cloud droplet spectrum as the true values, evaluating and comparing the predicted values ​​of the microphysical quantities of the machine learning model, and the evaluation methods include one or more of the mean square error (MSE), the mean absolute error (MAE) and the root mean square error (RMSE).

[0041] A computer program product is also provided, wherein the storage medium stores an executable computer program, and when the computer program is executed by a processor, it can implement the steps of the above-mentioned method for predicting microphysical quantities after the entrainment mixing and evaporation process in the cloud or the steps of the above-mentioned machine learning model training method for predicting microphysical quantities after the entrainment mixing and evaporation process in the cloud.

[0042] Beneficial effects: Traditional parameterization schemes require the establishment of specific expression formulas. Since the entrainment and mixing process in the cloud occurs at a small scale and is difficult to observe directly, and there are many influencing factors in the entrainment and mixing process, and the modes of action are complicated, it is difficult to obtain relatively accurate microphysical quantities after entrainment and mixing through traditional parameterization formulas. The microphysical quantity prediction method of this embodiment uses physical quantities that can be obtained relatively accurately in weather / climate models and have a high degree of influence on the entrainment and mixing process as characteristic quantities, and uses the microphysical quantities after the entrainment and mixing evaporation process in the cloud - number concentration Nc and discreteness dis as prediction quantities to train the machine learning model, and uses the trained machine learning model to predict the microphysical quantities after the entrainment and mixing evaporation process in the cloud. Compared with the traditional parameterization scheme, this method does not need to establish a specific physical relationship between the characteristic quantity and the prediction quantity through assumptions, and can obtain more accurate microphysical quantities after the entrainment and mixing evaporation process. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a schematic flow chart of a method for predicting microphysical quantities after the cloud entrainment mixing evaporation process according to the present invention;

[0044] Figure 2 It is a parameterized flow diagram of a method for predicting microphysical quantities after the cloud entrainment mixing evaporation process according to the present invention;

[0045] Figure 3 It is a schematic diagram of the training of a machine learning model for the method for predicting microphysical quantities after the cloud entrainment mixing evaporation process of the present invention;

[0046] Figure 4 is the feature importance ranking of log concentration (Nc);

[0047] Figure 5 Ranking the importance of features for dispersion (dis);

[0048] Figure 6 Comparison diagrams of (a-c) the number concentrations (Nc(P1), Nc(P2), Nc(P3)) obtained using the traditional parameterization scheme and (d) the number concentration (Nc(LGB)) predicted by the LGB model and the number concentration (Nc) calculated from the original cloud droplet spectrum;

[0049] Figure 7 Comparison of (a) the discreteness (dis(P3)) obtained using the traditional parameterization scheme and (b) the number concentration (dis(LGB)) predicted by the LGB model with the discreteness (dis) calculated from the original cloud droplet spectrum. DETAILED DESCRIPTION

[0050] The present invention is further described in detail below in conjunction with specific implementation methods.

[0051] The weather forecasting system of this embodiment includes a processor and a computer program product. The processor runs the computer program to pre-train a machine learning model with microphysical quantity prediction capabilities and executes the following steps: Figure 1 The steps of the method for predicting microphysical quantities after the entrainment mixing and evaporation process in the cloud are as follows: a parameterized data set of the entrainment mixing and evaporation process in the cloud is obtained based on the explicit mixing bubble model EMPM, the parameterized data set is input into a pre-trained machine learning model, and the corresponding number concentration Nc and discreteness dis output by the machine learning model are used as the predicted microphysical quantities after the entrainment mixing and evaporation process in the cloud.

[0052] The following uses an example to illustrate the specific process of the processor executing the method.

[0053] (1) The parameters of the entrainment mixing process are extracted by the explicit mixing bubble model EMPM, thereby obtaining a parameterized data set of the entrainment mixing evaporation process in the cloud. The explicit mixing bubble model EMPM is an existing technology designed for studying the entrainment mixing process. The model includes three processes: first, the cloud begins adiabatic condensation growth movement and stops rising after reaching the entrainment height; second, the ambient air block is entrained into the cloud under the action of turbulent vorticity, and at the same time, the air blocks in the cloud of the same size are entrained out of the cloud; third, the unsaturated dry air blocks entrained in the cloud are broken into air blocks of different sizes under the action of turbulence and randomly distributed throughout the cloud. Each cloud droplet undergoes a corresponding growth or evaporation process according to the supersaturation of its surrounding environment. The data of this embodiment comes from a large number of simulations of the entrainment mixing process in the cloud by the explicit mixing bubble model EMPM. These simulations take into account different factors that affect the entrainment mixing mechanism and are universal to the entrainment mixing evaporation process in the cloud that is not limited to a specific area. This embodiment uses the above-mentioned explicit mixed bubble model EMPM to extract the parameters of the physical process in the weather / climate model to form a parameterized data set of the entrainment mixing evaporation process in the cloud.

[0054] See Figure 2 ,This embodiment uses meteorological element observation data to drive EMPM, and uses different parameters of the above three processes as input combinations, simulates the cloud entrainment and mixing process tens of thousand times, and uses 10 different random seed simulations for the more than 10,000 simulation processes that have not dissipated.

[0055] This embodiment obtains a parameterized data set of the cloud entrainment mixing evaporation process based on the explicit mixing bubble model. Specifically, the model simulation area and its initial parameter settings are obtained. The initial parameters include the model scale Dm, vertical velocity w, adiabatic number concentration Na, and the ratio of entrained ambient air L. fraction, ambient air relative humidity RHe and turbulent dissipation rate ε; multiple groups of sensitivity tests are carried out based on the initial parameters to obtain the parameter changes of the cloud entrainment mixing evaporation process, and the parameters before and after the cloud entrainment mixing evaporation process are recorded in the parameterized data set respectively.

[0056] For the parameterized data set obtained from the simulation and sensitivity test results based on EMPM, the number concentration Nc and the dispersion dis after the entrainment mixed evaporation process are calculated by the traditional parameterization scheme (these two parameters will be used as the prediction quantities of the machine learning model LightGBM in the following text of this embodiment). The number concentration Nc and the dispersion dis of all parameterized data sets are clustered to obtain multiple subsets. The influence value of each parameter in each subset of the parameterized data set on the number concentration Nc and the dispersion dis is calculated by the SHAP model. The sum of the influence values ​​of all subsets is taken and sorted to characterize the influence of all parameters in the parameterized data set. The influence degrees are clustered, and all parameters in the first two groups are selected as the characteristic quantities of the machine learning model LightGBM in the following text of this embodiment. The selected characteristic quantities include microphysical quantities that are closely related to the entrainment mixed evaporation process and in weather or climate patterns: microphysical quantities inside the cloud and environmental quantities outside the cloud. Among them, the microphysical quantities inside the cloud include: adiabatic water content (LWCa), adiabatic number concentration (Na), adiabatic volume mean radius (r va ), the water content after entrainment, mixing and evaporation (LWCc); the extra-cloud environmental quantities include: turbulent dissipation rate (ε), relative humidity of the ambient air (RHe), water vapor mixing ratio of the ambient air (qve), temperature of the ambient air (Te) and air pressure of the ambient air (Pe).

[0057] SHAP explains the marginal contribution of input variables to model prediction from a global perspective, as well as the reason why the model makes predictions on instance samples. The additive explanatory model trained with SHAP is a generalized weighted linear model that fits the classifier to be explained. The prediction result of the model for any sample is expressed as the sum of the average prediction expectation of all samples and the SHAP values ​​of all features of the corresponding sample. This embodiment establishes an additive explanatory model based on SHAP, and regards all feature quantities as "contributors". For each predicted sample, the model generates a prediction value. The SHAP value represents the numerical value assigned to each feature quantity in the sample. The SHAP value of each feature quantity (Φj, the SHAP value of the jth feature value) represents the contribution of the feature quantity to the predicted quantity. Specifically, the parameter Ij, which is calculated by taking the average of the SHAP absolute values ​​of all samples of the feature quantity, is used to represent the influence of the feature quantity on the predicted quantity. The larger the value of parameter Ij, the greater the contribution of the feature quantity to the predicted value.

[0058] This embodiment uses SHAP to sort the parameters of multiple subsets of the parameterized data set, realizes the preliminary feature screening of the parameterized data set, and can compare the importance of the feature quantity after the subsequent machine learning model learning to further verify the rationality of feature selection. (2) Machine learning model LightGBM.

[0059] The machine learning model training method of this embodiment includes the following steps.

[0060] Obtain a parameterized data set, use the cloud microphysical quantities and cloud environment quantities in the parameterized data set as feature quantities, and use the number concentration Nc and the discreteness dis after the entrainment mixing evaporation process as prediction quantities; divide all feature quantities and prediction quantities into two groups, as a training set and a validation set, respectively. This embodiment uses the train_test_split function of the prior art to randomly divide the data set into a training set and a test set.

[0061] Train the machine learning model, use the feature quantity of the training set as the model input, and the prediction quantity as the model output to train the model, so as to obtain the ability to predict the microphysical quantities after the entrainment mixing evaporation process in the cloud through the input microphysical quantities in the cloud and the environmental quantities outside the cloud;

[0062] Model evaluation: The microphysical quantity prediction performance of the trained machine learning model is evaluated using the validation set and compared with the traditional parameterized microphysical quantity prediction results.

[0063] Machine learning has a significant effect on extracting complex nonlinear relationships from data, see Figure 3 In this embodiment, the above-mentioned microphysical quantities obtained by the explicit mixing bubble mode EMPM are used as characteristic quantities, and the number concentration Nc and discreteness dis after the entrainment mixing evaporation process are used as prediction quantities to train the machine learning model, so as to realize the ability to predict the microphysical quantities after the entrainment mixing evaporation process in the cloud by inputting the microphysical quantities in the cloud and the environmental quantities outside the cloud.

[0064] Existing machine learning models include extreme learning machine (ELM), multi-label learning (MLL), support vector machine (SVM), decision tree (DT), random forest (RF), gradient boosting decision tree (GBDT), etc. The machine learning model of this embodiment adopts Light Gradient Boosting Machine (LightGBM), which is an improved algorithm based on GBDT under the Boosting framework. It is a gradient boosting machine learning framework with decision tree as the base learning period, which has the advantages of high training efficiency, low memory usage, high accuracy, support for parallel learning, and rapid processing of massive data.

[0065] In the machine learning model LightGBM of this embodiment, the relative influence and importance of the feature quantity on the predicted quantity within a preset range are analyzed. SHAP is to explain the marginal contribution of the input variable to the model prediction from a global perspective, as well as the reason why the model predicts the instance sample. The additive explanation model trained with SHAP is a generalized weighted linear model to fit the classifier to be explained. The prediction result of the model for any sample is expressed as the sum of the average prediction expectation of all samples and the SHAP value of all features of the corresponding sample. This embodiment establishes an additive explanation model based on SHAP, and regards all feature quantities as "contributors". For each predicted sample, the model generates a prediction value. The SHAP value represents the value assigned to each feature quantity in the sample. The SHAP value of each feature quantity (Φj, the SHAP value of the jth feature value) represents the contribution of the feature quantity to the predicted quantity. Specifically, the parameter Ij obtained by taking the average SHAP absolute value of all samples of the feature quantity is used to characterize the influence of the feature quantity on the predicted quantity. The larger the value of the parameter Ij, the greater the contribution of the feature quantity to the predicted value.

[0066] The parameters Ij and ranking of the characteristic quantities of number concentration Nc are shown in Figure 4 (a), the horizontal axis is the absolute average value of the feature quantity SHAP, that is, the size of the parameter Ij. In the prediction model of logarithmic concentration Nc, the top three features with the greatest effect are Na, LWCc and r va Among them, Na contributes the most, and its SHAP value accounts for 41.1% of the total SHAP value of all feature quantities.

[0067] See Figure 4(b) is a feature density scatter plot. Each point in the figure represents a sample. The dispersion change indicates the high and low of the feature value. The redder the color, the larger the value of the feature itself, and the bluer the color, the smaller the value of the feature itself. The overlapping points jitter in the Y-axis direction, and the wide area indicates that there are a large number of samples gathered. The SHAP value distribution of each feature quantity is intuitively represented by the diagram. Na is positively correlated with Nc. Na represents the adiabatic number concentration, that is, the larger the adiabatic number concentration, the larger the number concentration after entrainment evaporation. Secondly, LWCc is positively correlated with Nc. LWCc can characterize the intensity of entrainment mixed evaporation to a certain extent. The larger the LWCc, the weaker the entrainment evaporation intensity, and the smaller the LWCc, the stronger the entrainment evaporation intensity. Therefore, the positive correlation between LWCc and Nc indicates that the stronger the entrainment mixed evaporation intensity, the smaller the Nc, that is, the more the number concentration decreases. Again, r va Negatively correlated with Nc. When LWCa is constant, the larger Na is, the higher r va The smaller the rva, the negatively correlated with Nc. In addition, RHe, Pe, qve, and Te have little effect on Nc, and their correlation with Nc is unclear.

[0068] Correspondingly, the dispersion dis Figure 5 , the characteristic quantities that have a greater impact on it are LWCc, ε and RHe. LWCc is negatively correlated with dis, that is, the greater the entrainment mixing evaporation intensity, the greater the dis, and the more the cloud droplet spectrum broadens. RH and ε are both negatively correlated with dis. The smaller the RHe, the smaller the ε, and the more non-uniform the corresponding entrainment mechanism. Therefore, the correlation between RHe and ε and dis indicates that the more non-uniform the entrainment mechanism, the more the cloud droplet spectrum broadens, which is consistent with the theoretical study that the cloud droplet spectrum broadening caused by non-uniform entrainment mixing is stronger than that of the uniform mixing mechanism.

[0069] (3) Analysis of the predictive validity of this embodiment.

[0070] In this embodiment, 80% of the data in the data set consisting of feature quantities and prediction quantities are randomly selected to train LightGBM, and the remaining 20% ​​of the data are used to compare and evaluate the prediction performance of the model with the traditional parameterization scheme.

[0071] The uniform entrainment and mixing parameterization assumed in the current weather model is defined as P1, the extreme non-uniform entrainment and mixing parameterization assumed in the climate model is defined as P2, and the parameterization for real-time diagnosis of the current entrainment and mixing mechanism based on the cloud droplet properties and environmental conditions at each moment is defined as P3. Figure 6 (a) to (c) show the comparison between the number concentration after entrainment and mixing obtained using three traditional parameterization schemes and the number concentration calculated based on the original cloud droplet spectrum. Figure 6 (d) shows a comparison chart between the number concentration predicted by the LightGBM model obtained after training with the number concentration calculated based on the original cloud droplet spectrum.

[0072] Since more data points fall on the diagonal line of the figure, it means that the parameterization or machine learning model has a good prediction effect. It can be seen that Figure 6 For the prediction of number concentration after entrainment mixing, the trained LGB machine learning model is much more accurate than the traditional parameterization scheme. Figure 7 The prediction effects of the traditional parameterization scheme (a) and the LGB model (b) on the dispersion after the entrainment mixing process are given. Similarly, compared with the traditional parameterization scheme, the machine learning model can obtain more accurate dispersion.

[0073] In this embodiment, the microphysical quantities calculated based on the cloud droplet spectrum are used as true values ​​for reference comparison and evaluation. The evaluation methods are mean square error (MSE), mean absolute error (MAE) and root mean square error (RMSE):

[0074]

[0075] Among them, y is the true value, y' is the predicted value obtained by parameterization or machine learning methods, and m is the number of data.

[0076] See Table 1 below, the MAE, MSE and RMSE of the traditional parameterization method and the LGB machine learning model for predicting the microphysical quantities (number concentration Nc and discreteness dis) after entrainment and mixing. From the numerical comparison of the loss function, it can also be seen that the effect of the LGB machine learning model trained in this project is far better than the traditional parameterization scheme.

[0077] Table 1. MAE, MSE, and RMSE for predicting the number concentration (Nc) and dispersion (dis) after the entrainment mixing process using the traditional parameterization scheme and the LGB machine learning model.

[0078]

[0079] The method for predicting microphysical quantities after the entrainment, mixing and evaporation process in clouds in this embodiment utilizes the advantage of machine learning methods that do not require the establishment of specific physical relationship equations between characteristic quantities and predicted quantities. Physical quantities that can be obtained in weather / climate models and are crucial to the entrainment and mixing process are selected as characteristic quantities, and number concentrations and discreteness after the entrainment and mixing process are used as predicted quantities. The machine learning model is trained. Compared with traditional parameterization schemes, this method can obtain more accurate microphysical quantities after the entrainment, mixing and evaporation process.

[0080] The above is only an implementation method of the invention, and does not limit the scope of patent protection. Those skilled in the art can make non-substantial changes or substitutions based on the invention, which still fall within the scope of patent protection.

Claims

1. A method for predicting microphysical quantities after cloud entrainment, mixing and evaporation, characterized in that: The steps include: Based on the explicit mixing bubble model, a parameterized data set of the entrainment mixing evaporation process in the cloud is obtained. Specifically, the model simulation area and its initial parameter settings are obtained. The initial parameters include the model scale Dm, vertical velocity w, adiabatic number concentration Na, and the ratio of entrained ambient air L. fraction , ambient air relative humidity RHe and turbulent dissipation rate ε; based on the initial parameters, multiple groups of sensitivity tests were carried out to obtain the parameter changes of the cloud entrainment mixing evaporation process, and the parameters before and after the cloud entrainment mixing evaporation process were recorded in the parameterized data set respectively; The number concentration Nc and the dispersion dis after the entrainment mixed evaporation process are calculated by the above parameterized data set, and the number concentration Nc and the dispersion dis of all parameterized data sets are clustered to obtain multiple subsets, and the influence value of each parameter in these subsets on the number concentration Nc and the dispersion dis is calculated. The sum of the influence values ​​of each parameter in all subsets is taken and sorted to characterize the influence of each parameter in the parameterized data set, and the influence is clustered, and the parameters with the influence reaching the preset degree are selected as the feature quantity of the machine learning model; The characteristic quantities screened from the parameterized data set include microphysical quantities within the cloud and environmental quantities outside the cloud; The parameterized data set is input into a pre-trained machine learning model, and the corresponding number concentration Nc and discreteness dis output by the machine learning model are used as the predicted microphysical quantities after the entrainment mixing and evaporation process in the cloud.

2. The method for predicting microphysical quantities after the cloud entrainment mixing and evaporation process according to claim 1 is characterized in that: The cloud microphysical quantities include: adiabatic water content LWCa, adiabatic number concentration Na, adiabatic volume average radius r va And multiple of the water content LWCc after entrainment and mixing evaporation.

3. The method for predicting microphysical quantities after the cloud entrainment mixing and evaporation process according to claim 1 is characterized in that: The cloud-outside environmental quantities include: multiple ones of: turbulence dissipation rate ε, ambient air temperature Te, ambient air relative humidity RHe and ambient air water vapor mixing ratio qve.

4. A computer program product having an executable computer program stored thereon, characterized in that: When the computer program is executed, the steps of the method for predicting microphysical quantities after the entrainment, mixing and evaporation process in a cloud as described in any one of claims 1 to 3 are implemented.

5. A machine learning model training method for predicting microphysical quantities after entrainment, mixing and evaporation in clouds, characterized by: The steps include: Obtain a parameterized data set, use the cloud microphysical quantities and cloud environment quantities processed as claimed in any one of claims 1 to 4 in the parameterized data set as characteristic quantities, and use the number concentration Nc and the dispersion dis after the entrainment mixing evaporation process as prediction quantities; divide all characteristic quantities and prediction quantities into two groups, which are respectively used as a training set and a validation set; Train the machine learning model, use the feature quantity of the training set as the model input, and the prediction quantity as the model output to train the model, so as to obtain the ability to predict the microphysical quantities after the entrainment mixing evaporation process in the cloud through the input microphysical quantities in the cloud and the environmental quantities outside the cloud; Model evaluation: Use the validation set to evaluate the microphysical quantity prediction performance of the trained machine learning model, that is, compare the microphysical quantities predicted by the machine learning model with the parameterized results of the microphysical quantities.

6. The machine learning model training method for predicting microphysical quantities after the cloud entrainment mixing and evaporation process as claimed in claim 5 is characterized by: The machine learning models include: extreme learning machine ELM, multi-label learning MLL, support vector machine SVM, decision tree DT, random forest RF, gradient boosting tree GBDT or lightweight gradient boosting machine LightGBM.

7. The machine learning model training method for predicting microphysical quantities after the cloud entrainment mixing and evaporation process as claimed in claim 5 is characterized by: An additive explanatory model based on SHAP is established. For each feature quantity of the parameterized data set, the additive explanatory model generates a prediction value for each prediction sample based on all feature quantities, and the SHAP value of each feature quantity is used to characterize the contribution of the feature quantity to the prediction quantity. The larger the absolute value of the SHAP value, the greater the contribution to the prediction value.

8. The machine learning model training method for predicting microphysical quantities after the cloud entrainment mixing and evaporation process as claimed in claim 5 is characterized by: The model evaluation is specifically as follows: taking the microphysical quantities calculated based on the cloud droplet spectrum as the true values, evaluating and comparing the predicted values ​​of the microphysical quantities of the machine learning model, and the evaluation methods include one or more of the mean square error (MSE), the mean absolute error (MAE) and the root mean square error (RMSE).

9. A computer program product having an executable computer program stored thereon, characterized in that: When the computer program is executed, it implements the machine learning model training method for predicting microphysical quantities after the entrainment mixing evaporation process in the cloud as described in any one of claims 5 to 8.

Images