A method and system for accurately predicting CCN number concentration with high temporal and spatial resolution
By combining WRF-Chem numerical mode and machine learning technology, a high-spatial-time resolution CCN number concentration accurate prediction system was established, which solved the spatial-temporal resolution and accuracy of CCN number concentration prediction in the existing technology, and achieved higher prediction accuracy and general Adaptability, helps to assess the impact of CCN on weather and climate.
Patent Information
- Application Number
- CN202410012958.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-04
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-01-04
AI Technical Summary
The existing direct observation of CCN number concentrations and satellite remote sensing technologies are difficult to achieve high spatial and temporal resolution and accurate prediction, especially in polluted areas, and the existing model lacks universality for the parameterization scheme of aerosol-cloud.
Combining WRF-Chem numerical mode and machine learning technology, a high-temporal and spatial resolution CCN number concentration accurate prediction system based on 'observation-mode-intelligent calculation' is established. By collecting CCN number concentration samples and characteristic factors, pre-processing, screening and model construction , a nonlinear relationship model is constructed using a random forest algorithm to predict CCN number concentration.
The prediction accuracy of CCN number concentration is improved, the shortcomings of the space-time scale are made up for, and the accurate evaluation of CCN number concentration is achieved, which is universal and can more accurately evaluate its impact on weather and climate.
Smart Images

Figure CN117828992B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of atmospheric measurement technology, and in particular to a method and system for accurately predicting CCN number concentration with high temporal and spatial resolution. Background Art
[0002] The impact of atmospheric aerosols on cloud radiative forcing is the largest source of uncertainty in current climate change predictions. The main reason for this uncertainty is that atmospheric aerosols can act as cloud condensation nuclei (CCN). CCN refers to atmospheric aerosol particles that can be activated to form cloud droplets under supersaturated conditions. The interaction between CCN and water vapor affects the micro- and macroscopic physics of clouds, thereby affecting cloud formation, properties, dynamics and precipitation, further affecting the Earth's energy budget, and thus affecting climate and weather. Therefore, conducting relevant research on atmospheric aerosols and CCN is of great scientific significance and is crucial to improving the aerosol-cloud parameterization scheme in climate models to accurately assess the indirect climate effects of aerosols.
[0003] The CCN number concentration is related to the environmental supersaturation. The greater the supersaturation, the greater the CCN number concentration, and then there will be more aerosol particles with cloud-forming potential. Under given supersaturation conditions, the aerosol particle spectrum distribution, chemical composition and mixing state play an important role in determining CCN activity. In order to quantify the impact of CCN on the Earth's radiation budget, the key is to obtain the CCN number concentration on a spatiotemporal scale. However, the characteristics of aerosols will show spatiotemporal differences with the physical and chemical processes of particles in the atmosphere. Generally, the closer the area is to the emission source, the higher the concentration of atmospheric gaseous precursors, oxidative free radicals, and primary emission particles (such as primary organic matter and BC from biomass, etc.), which will make various atmospheric chemical and physical processes (particle nucleation, condensation growth, collision growth, homogeneous and heterogeneous reactions, etc.) more intense and rapid. These processes may change the size and composition of aerosol particles on a shorter time scale, thereby changing their hygroscopic and activation properties. The high spatiotemporal differences in the characteristics of aerosol particles make the CCN number concentration (N CCN ) also have significant differences in their spatial and temporal distribution, which highlights the importance of quantifying N in different regions. CCN The importance of.
[0004] Existing direct observations of CCN are mostly based on short-term field observations at some single sites, and current satellite remote sensing technology cannot achieve direct detection of CCN. Although ground-based and air-based remote sensing observations can be used to obtain aerosol optical thickness characteristics on a global scale to invert CCN number concentrations, the satellite optical method is more sensitive to particles in large particle sizes, resulting in large errors in the accuracy of its inversion. In addition, the current model still has great uncertainty in predicting CCN number concentrations in polluted areas, and it is still extremely challenging to obtain accurate values of CCN number concentrations over a large spatial scale. Closed-loop experimental studies of CCN help to clarify the mechanism of the conversion of aerosol particles into CCN, as well as the relationship between other atmospheric variables and parameters and CCN, but due to the limitations of observational data and the nonlinear relationship between CCN number concentration and these factors, a large-scale universal CCN parameterization scheme cannot yet be obtained. Summary of the invention
[0005] In order to solve the technical problems in the above background, the present invention combines WRF-Chem numerical model simulation and machine learning technology to establish a high temporal and spatial resolution CCN number concentration accurate prediction system based on "observation-model-intelligent measurement". This method realizes the correction of current climate model results and helps to clarify the refined temporal and spatial variation characteristics of CCN number concentration.
[0006] To achieve the above object, the present invention provides a method for accurately predicting CCN number concentration with high temporal and spatial resolution, the steps comprising:
[0007] Use the WRF-Chem model to collect CCN number concentration samples that affect the area to be predicted; use a multi-source big data network to collect characteristic factors of the area to be predicted;
[0008] Preprocessing the CCN number concentration sample and the characteristic factor respectively;
[0009] Screening the processed characteristic factors to obtain prediction factors;
[0010] Based on the prediction factors, a prediction model is constructed, and the prediction model is used to complete the prediction of CCN number concentration.
[0011] Preferably, the characteristic factors include: chemical composition data, conventional gas pollutant data and meteorological element data set.
[0012] Preferably, the method for performing the preprocessing includes: normalizing the characteristic factors; and simultaneously performing memory correction and generalization processing on the CCN number concentration data to obtain the processed data.
[0013] Preferably, the processed characteristic factors are screened to eliminate characteristic factors with little influence on the model output results to obtain the prediction factors.
[0014] Preferably, the method for constructing the prediction model includes: establishing a nonlinear relationship between the characteristic factor and the CCN number concentration data; wherein the CCN number concentration data is a target variable.
[0015] Preferably, the prediction model is constructed using a random forest algorithm; meanwhile, Cartesian coordinates are added during the construction process.
[0016] The present invention also provides a high temporal and spatial resolution CCN number concentration accurate prediction system, the system is used to implement the above method, including: an acquisition module, a preprocessing module, a screening module and a prediction module;
[0017] The acquisition module is used to collect CCN number concentration samples affecting the area to be predicted using the WRF-Chem model; and to collect characteristic factors of the area to be predicted using a multi-source big data network;
[0018] The preprocessing module is used to preprocess the CCN number concentration sample and the characteristic factor respectively;
[0019] The screening module is used to screen the processed characteristic factors to obtain prediction factors;
[0020] The prediction module is used to construct a prediction model based on the prediction factors, and use the prediction model to complete the prediction of CCN number concentration.
[0021] Preferably, the characteristic factors include: chemical composition data, conventional gas pollutant data and meteorological element data set.
[0022] Compared with the prior art, the present invention has the following beneficial effects:
[0023] The CCN prediction system constructed by the present invention can optimize the model simulation results and improve the prediction accuracy. Compared with the traditional optical property inversion algorithm, the predicted CCN number concentration has higher accuracy. At the same time, it also makes up for the shortcomings of CCN observation in the temporal and spatial scales. At the same time, the prediction method of the present invention is universal, and its accurate prediction of CCN number concentration helps to parameterize its temporal and spatial characteristics into the model simulation, thereby accurately evaluating its impact on weather and climate. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0025] Figure 1 A schematic diagram of a method flow chart of an embodiment of the present invention;
[0026] Figure 2 A schematic diagram of a model architecture of an embodiment of the present invention;
[0027] Figure 3 This is a schematic diagram of the importance ranking of characteristic factors of the CCN number concentration prediction model according to an embodiment of the present invention;
[0028] Figure 4 Schematic diagram of determining the hyperparameters of the CCN number concentration prediction model in an embodiment of the present invention; wherein (a) the cross-validation score changes with the number of trees, (b) the cross-validation score changes with the maximum number of features, (c) the cross-validation score changes with the minimum sample size of leaf nodes, (d) the cross-validation score changes with the minimum number of node segmentation samples, and (e) the cross-validation score changes with the depth of the decision tree. DETAILED DESCRIPTION
[0029] 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.
[0030] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0031] Embodiment 1
[0032] like Figure 1 FIG. 1 is a schematic diagram of the method flow of this embodiment, and the steps include:
[0033] S1. Use the WRF-Chem model to collect CCN number concentration samples that affect the area to be predicted; use a multi-source big data network to collect characteristic factors of the area to be predicted;
[0034] Among them, the characteristic factors that affect the target variables include: chemical composition data (using Tsinghua University TAP data set), conventional gas pollutant data (using national control station data), and meteorological element data (using ERA-5 reanalysis data).
[0035] The simulation of the target variable CCN number concentration data adopts the regional online coupled meteorological and chemical model (The Weather Research and Forecasting Model / Chemistry, WRF-Chem). The meteorological data used in the numerical model are the FNL reanalysis data released in real time by the Numerical Weather Prediction Center (NWP) of the National Oceanic and Atmospheric Administration (NOAA) of the United States; the emission source data list mainly includes anthropogenic source list, natural source (biogenic source), and biomass burning emission data. The anthropogenic source list comes from the Multi-resolution Emission Inventory for China (MEIC) model. This gridded inventory data set was developed by Tsinghua University. The temporal and spatial resolutions are monthly and 0.25°*0.25° respectively. Its data range scale can cover the entire Chinese region. Its polluting industries involve industry, electricity, residents, transportation, and agriculture, covering more than 700 anthropogenic emission sources, and pollutants include BC, OC, PM2.5, PMcoarse, CO, CO2, NH3, NOx, SO2, and VOC. The biogenic source list comes from the results of the biological emission source model of MEGAN (Model of emissions of gases and aerosols from Nature). The biogenic sources mainly include organic matter such as isoprene and terpenes emitted by plants. Different microphysical schemes and chemical mechanisms are selected to comprehensively simulate and predict the spatiotemporal variation of CCN number concentration over a long time scale in the study area, comprehensively analyze the error, predict the duration and the complexity of the model setting, and select the optimal model to output the results.
[0036] S2. Preprocessing the CCN number concentration sample and the characteristic factor respectively;
[0037] The sample data are matched in time and space to ensure the uniformity of the time and space resolution of the sample data; the near-surface CCN number concentration data output by the model is selected (taking the typical supersaturation S=0.2% in the cloud as an example); the variables of the characteristic factors are normalized, and all the data of the characteristic factor columns are converted to [0, 1], in order to eliminate the error caused by the order of magnitude difference between the data of each dimension to the model prediction; in the embodiment, whether the characteristic factor contains chemical component information is considered respectively to evaluate the model prediction performance.
[0038] S3. Screen the processed characteristic factors to obtain prediction factors.
[0039] This embodiment combines the characteristic parameter importance evaluation to screen the input characteristic factors, and eliminates the characteristic factors that have little impact on the output results, such as chloride salt ChL in chemical components, sea level pressure MSL and total cloud cover TCC in meteorological factors. The importance of characteristic parameters is evaluated by the method of rearranging the importance of characteristics. The main principle is to randomly arrange all the values of a certain feature in the data set and recalculate the prediction results. If the difference between the two prediction results is large, it means that the characteristic factor has a significant impact on the model prediction results. If the difference before and after is not large, it means that the importance of the characteristic parameter is low, so as to obtain the prediction factor.
[0040] S4. Based on the prediction factors, a prediction model is constructed, and the prediction model is used to complete the prediction of CCN number concentration.
[0041] The specific model construction ideas are as follows:
[0042] Constructing a machine learning model means establishing a nonlinear relationship between the target variable and the characteristic parameters. The target variable of this embodiment is to use the WRF-Chem numerical model, select different microphysical schemes and chemical mechanisms, conduct comprehensive simulation and prediction of the spatiotemporal variation of CCN number concentration over a long time scale in the study area, conduct comprehensive error analysis, predict the duration and complexity of the model setting, and select the optimal model to output the results. The overall model architecture is as follows: Figure 2 shown.
[0043] On the basis of the WRF-Chem numerical model, combined with the more common atmospheric state variables, the present invention selects PM2.5 mass concentration and chemical component data (organic matter OM, ammonium salt NH4, sulfate SO4, nitrate NO3, black carbon BC), conventional gas pollutant data (ozone O3, sulfur dioxide SO2, nitrogen dioxide NO2, carbon monoxide CO), meteorological data from the European Center for Reanalysis Meteorological Data (temperature TEM, relative humidity RH, wind speed WS, wind direction WD, boundary layer height BLH, total precipitation PRE, surface pressure SP, solar radiation SSR) and other conventional atmospheric data as input features. The random forest model of the machine learning integrated algorithm is used to construct the CCN number concentration prediction method, and a comparative experiment is conducted with the established numerical model and actual observation results, and the optimal ensemble prediction scheme is found by comprehensively considering the prediction time, model complexity and error analysis.
[0044] Since the information of time and space scale is involved, Cartesian coordinates are added to the characteristic factors at the beginning of model construction. Since the time information DOY (expressed as the day of the year) cannot reflect the seasonality of time, because the first and last days of the year are both in winter, but the DOY values are very different. In order to solve this problem, the number of days in a year DOY is converted into a Cartesian coordinate system:
[0045]
[0046] Among them, t x , t y is the Cartesian coordinate system; T is the total number of days in a year; based on formula (1), DOY can be normalized to the interval [02π] and further converted into polar coordinates. Spatial information is generally given in polar coordinates of longitude and latitude. Since the polar coordinates of longitude and latitude vary unevenly, they are not suitable for distance representation. The Cartesian coordinate system can achieve uniform numerical variation. Therefore, the longitude and latitude information of the national control station is converted into the Cartesian coordinate system as the characteristic parameter input model. The specific conversion formula is as follows:
[0047]
[0048] in, is the latitude, θ is the longitude, and R is the radius of the earth.
[0049] The adjustment of model hyperparameters is mainly to limit model parameters, so that the complexity of the model is reduced while the model output results are better. The hyperparameter adjustments involved here mainly include the number of trees, the maximum growth depth of the tree, the minimum sample size of leaves, the minimum sample size of branch nodes, the maximum number of characteristic factors and standards, and the Gini index is generally used by default. A model that is too simple or too complex will result in poor generalization ability of the model, so it is necessary to seek the optimal model complexity to minimize the generalization error. Based on the idea of cross_val_score cross-validation, that is, setting different random forest parameters, first divide the training set into n parts, where n is 10 parts, use n-1 parts to train the model, and use the remaining 1 part to predict the model. The optimal parameters of the model are obtained based on the average of the 10-fold cross-validation results. The larger the number of trees, the more complex the model is and the better the predictive ability of the model is. The maximum growth depth of the tree does not need to be limited by default, but it needs to be limited when the sample size is large. Limiting this parameter simplifies the model. The minimum number of samples of leaves and the minimum number of samples of branch nodes are both pruning operations. Similarly, these two parameters need to be adjusted when the sample size is large. The maximum number of selected features, that is, how many features are selected from all input features to establish a decision tree, can be the default value or a directly specified number. If it is set to None, the maximum number of features is selected.
[0050] Embodiment 2
[0051] The following will explain in detail how the present invention solves technical problems in real life in conjunction with this embodiment.
[0052] Taking the North China Plain as an example, the WRF-Chem model was used to simulate aerosols using two mechanisms, the carbon bond mechanism-particle size method (CBMZ-MOSAIC) and the regional acid deposition mechanism-modal method (RADM2-SORGAM). The numerical model calculation efficiency was comprehensively compared with the six field observation results of the North China Plain in the early stage, and the optimal model simulation result was selected for output. The selection of characteristic variables refers to the discussion on the influence of previous field observations on CCN number concentration, such as the mass concentration of particulate matter that characterizes the degree of pollution, the meteorological factors that affect the secondary process of particulate matter, the concentration of gaseous pollutants that affect the nucleation rate of particulate matter, and the chemical composition information of particulate matter. Therefore, at the initial stage of model construction, PM2.5 mass concentration and chemical composition data (organic matter, ammonium salt, sulfate, nitrate, black carbon) were selected, conventional gas pollutant data came from observations at national control stations (ozone, sulfur dioxide, nitrogen dioxide, carbon monoxide), and meteorological data came from European Center reanalysis meteorological data (temperature, relative humidity, wind speed, wind direction, boundary layer height, total precipitation, surface pressure, solar radiation) and other conventional atmospheric data as input features. The model results were verified with the help of field observations in the North China Plain, which were observations at Beijing urban stations in the autumn and winter of 2014, the autumn of 2015, the autumn and winter of 2016, and the summer of 2017, observations at Xingtai, Hebei in the spring and summer of 2016, and observations at Gucheng, Hebei in the winter of 2018.
[0053] Afterwards, all data are corrected and assimilated to ensure uniform time resolution; the characteristic factors are normalized; in the embodiment, whether the characteristic factors take into account chemical component information or not is considered to evaluate the prediction performance of the model.
[0054] After the data processing is completed, the prediction model factors are interpreted. The model input prediction factors are screened according to the model ranking importance, and the characteristic factors with little influence on the model output results are screened out to reduce the model complexity; such as chloride salt ChL in chemical components, sea level pressure MSL and total cloud cover TCC in meteorological factors, etc. Figure 3 A graph showing the importance ranking of the feature factors selected for the model.
[0055] Finally, the model was constructed. Since it involved information on the spatiotemporal scale, Cartesian coordinates were added when constructing the model. The random forest algorithm of machine learning was determined to build the model, and the model hyperparameters were adjusted. By combining the model characteristic factors and target variables, a high spatiotemporal resolution CCN number concentration accurate prediction method based on "observation-model-machine learning" was established. Figure 4The optimization results of the hyperparameters of the random forest model constructed for the North China region. The CV cross-validation results show that when the number of trees (n_estimators) is less than 200, the prediction accuracy increases rapidly as the number of trees increases, and then gradually stabilizes. Previous studies have shown that the default number of trees in the random forest algorithm is 500. This embodiment selects the parameter to be 500 in combination with the CV score value and the data sample size. The influence of the maximum depth (max_depth) of the tree on the accuracy is manifested in that as the depth increases, the accuracy of the model shifts to the right, that is, the model complexity increases. The value here is 28. The larger the minimum sample number of leaves and the minimum sample number of branch nodes in the pruning operation, the greater the generalization error of the model, indicating that the model itself is at a level close to the optimal model complexity. Taking into account the large sample size, it is slightly increased here based on the default value. The influence of the selection of the maximum number of selected features on the CV score is manifested as a trend of first increasing and then decreasing, so the highest value of the CV curve is selected to correspond to 16. The comparison between the model test set of this embodiment and the WRF-Chem simulation value is: the correlation is 0.89-0.91, the slope is 0.83-0.86, and the root mean square error in the winter half year is slightly higher than that in the summer half year. Further comparison of the observed values of CCN number concentration, the WRF-Chem model simulation values, and the random forest model prediction values, taking the supersaturation of 0.20% as an example, the high spatiotemporal resolution CCN number concentration accurate prediction method based on "observation-model-machine learning" can achieve a better prediction of CCN number concentration in North China, which is about 30% higher than the WRF-Chem model simulation results, and the error of six observation comparisons and verifications is within about 20%. It shows that the constructed machine learning model can further realize the prediction of CCN number concentration in a larger spatiotemporal range in complex pollution areas.
[0056] It should be noted that in this embodiment, the prediction performance of the comparison model considering the chemical component information or not is that the prediction accuracy of the model in which the characteristic factor considers the component information is higher than that of the model that only includes the PM2.5 particle mass concentration and other atmospheric state variables.
[0057] In summary, based on the WRF-Chem model simulation, the combination of machine learning technology and conventional atmospheric state variables can achieve the prediction of cloud condensation nucleus number concentration on a larger temporal and spatial scale. Compared with the actual observation results, the machine learning model optimizes the WRF-Chem numerical model simulation results. The CCN number concentration prediction method proposed in this example is crucial to improve the aerosol-cloud parameterization scheme in the climate model to accurately evaluate the indirect climate effect of aerosols.
[0058] Embodiment 3
[0059] This embodiment also provides a high-temporal and spatial resolution CCN number concentration accurate prediction system, including: an acquisition module, a preprocessing module, a screening module and a prediction module; the acquisition module is used to collect CCN number concentration samples affecting the area to be predicted using the WRF-Chem model; the characteristic factors of the area to be predicted are collected using a multi-source big data network; the characteristic factors are characteristic factor data collected using a multi-source big data network; the preprocessing module is used to preprocess the CCN number concentration samples and characteristic factors respectively; the screening module is used to screen the processed characteristic factors to obtain prediction factors; the prediction module is used to build a prediction model based on the prediction factors, and use the prediction model to complete the prediction of CCN number concentration. Among them, the characteristic factors include: chemical component data, conventional gas pollutant data and meteorological element data sets.
[0060] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.
Claims
1. A method for accurately predicting CCN number concentration with high temporal and spatial resolution, characterized in that the steps include: The WRF-Chem model is used to collect CCN number concentration samples affecting the area to be predicted; Use multi-source big data network to collect characteristic factors of the area to be predicted; The characteristic factors include: chemical component data, conventional gas pollutant data and meteorological element data; Preprocessing the CCN number concentration sample and the characteristic factor respectively; The processed characteristic factors are screened to obtain prediction factors; specifically, the characteristic factors for input are screened in combination with the evaluation of the importance of characteristic parameters, and the characteristic factors with little influence on the output results are eliminated; the importance of characteristic parameters is evaluated by the method of re-arrangement of characteristic importance, the principle of which is to randomly arrange all values of a certain feature on the data set and recalculate the prediction results. If the difference between the two prediction results is large, it means that the characteristic factor has a significant influence on the prediction result of the model, and if the difference is small, it means that the importance of the characteristic parameter is low, thereby obtaining the prediction factor; The method for constructing a prediction model includes: establishing a nonlinear relationship between the characteristic factor and the CCN number concentration data; wherein the CCN number concentration data is a target variable; constructing the prediction model using a random forest algorithm; and adding Cartesian coordinates during the construction process; determining the random forest algorithm of machine learning for model construction, adjusting the model hyperparameters, and combining the model characteristic factors and the target variables to establish a high spatiotemporal resolution CCN number concentration accurate prediction method based on "observation-model-machine learning"; the method for adjusting the model hyperparameters includes: based on the idea of cross_val_score cross-validation, setting different random forest parameters, first dividing the training set into n parts, using n-1 parts to train the model, and using the remaining 1 part to predict the model, and obtaining the optimal parameters of the model based on the mean of the n-fold cross-validation results.
2. The method for accurately predicting CCN number concentration with high temporal and spatial resolution according to claim 1, characterized in that: The method for performing the preprocessing includes: normalizing the characteristic factors; and simultaneously performing memory correction and generalization processing on the CCN number concentration data to obtain the processed data.
3. The method for accurately predicting CCN number concentration with high temporal and spatial resolution according to claim 2 is characterized in that: The processed characteristic factors are screened, and characteristic factors with little influence on the model output result are eliminated to obtain the prediction factors.
4. A high temporal and spatial resolution CCN number concentration accurate prediction system, the system is used to implement the method described in any one of claims 1 to 3, characterized in that: include: Acquisition module, preprocessing module, screening module and prediction module; The acquisition module is used to collect CCN number concentration samples affecting the area to be predicted using the WRF-Chem model; and to collect characteristic factors of the area to be predicted using a multi-source big data network; The characteristic factors include: chemical component data, conventional gas pollutant data and meteorological element data; The preprocessing module is used to preprocess the CCN number concentration sample and the characteristic factor respectively; The screening module is used to screen the processed feature factors to obtain prediction factors; specifically, the feature factors for input are screened in combination with the feature parameter importance evaluation, and the feature factors with little influence on the output result are eliminated; the feature parameter importance is evaluated by the method of re-arrangement feature importance, the principle is to randomly arrange all values of a certain feature on the data set and recalculate the prediction result. If the difference between the two prediction results is large, it means that the feature factor has a significant impact on the model prediction result. If the difference is small, it means that the feature parameter is of low importance, thereby obtaining the prediction factor; The step of constructing the prediction model by the prediction module includes: establishing a nonlinear relationship between the characteristic factor and the CCN number concentration data; wherein the CCN number concentration data is the target variable; constructing the prediction model using a random forest algorithm; at the same time, adding Cartesian coordinates during the construction process; determining the random forest algorithm of machine learning for model construction, adjusting the model hyperparameters, combining the model characteristic factors and the target variables, and establishing a high spatiotemporal resolution CCN number concentration accurate prediction method based on "observation-model-machine learning"; the method of adjusting the model hyperparameters includes: based on the idea of cross_val_score cross-validation, setting different random forest parameters, first dividing the training set into n parts, using n-1 parts to train the model, using the remaining 1 part to predict the model, and obtaining the optimal parameters of the model based on the mean of the n-fold cross-validation results.
