A cloud-rain environment modeling and predicting method and device based on supervised machine learning
By using supervised machine learning methods and multi-source meteorological data to construct a cloud and rain environment modeling and prediction model, the problem of limited effectiveness of traditional methods at small and medium scales is solved, and refined modeling and prediction of cloud and rain environments are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINESE PEOPLES LIBERATION ARMY UNIT 61540
- Filing Date
- 2025-09-24
- Publication Date
- 2026-05-19
AI Technical Summary
Existing technologies are insufficient for achieving detailed modeling and prediction of cloud and rain environments. Traditional simulation methods have limited effectiveness at small and medium scales and cannot meet the needs of various applications.
A supervised machine learning-based approach was adopted to construct a cloud and rain environment modeling and prediction model using multi-source meteorological data. Through data collection, preprocessing, feature extraction, and model training, a cloud and rain field fusion prediction model was constructed by combining various meteorological parameters and variables.
It improves the accuracy of cloud and rain environment modeling and prediction, is applicable to a variety of situations, has good universality, and solves the problem of insufficient fine simulation at small and medium scales in traditional methods.
Smart Images
Figure CN121257294B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical fields of atmospheric environmental science and artificial intelligence deep learning, specifically to a method and apparatus for cloud and rain environment modeling and prediction based on supervised machine learning. Background Technology
[0002] Modeling research on cloud and rain environments not only helps to deepen the understanding of the mechanisms of cloud and rain formation, development and evolution, and improve the forecasting ability of complex weather phenomena such as precipitation, but also provides scientific support and decision-making reference for related fields. It is of great significance for equipment performance evaluation, aircraft safety assurance, and artificial weather modification, and has always been a research focus in related fields.
[0003] With the development of detection technologies, various observation and detection methods, such as micro-rain radar, raindrop spectrometers, fog droplet spectrometers, visibility meters, millimeter-wave radar, dual-polarization Doppler radar, wind profiler radar, microwave radiometers, active lightning location imaging systems, and ground electric field meters, detect the meteorological environment from different perspectives. Each of these devices has its strengths, and the data can only reflect a few meteorological elements, lacking a holistic understanding of the entire cloud and rain environment. Traditionally, cloud and rain environments are simulated using sub-grid scale parameter processes based on assimilation forecasting methods. However, the simulation effect for small- to medium-scale refined cloud and rain environments is limited and cannot meet the needs of various applications. It is necessary to explore the integration of traditional simulation data and historical statistical analysis features, combined with machine learning methods, to construct a refined cloud and rain environment model. Summary of the Invention
[0004] This invention primarily addresses the problem of improving the modeling and prediction accuracy of cloud and rain environments based on multi-source meteorological data. This invention discloses a method and apparatus for modeling and predicting cloud and rain environments based on supervised machine learning.
[0005] In a first aspect, this invention discloses a method for modeling and predicting cloud and rain environments based on supervised machine learning, comprising:
[0006] S1, Collect historical meteorological environment dataset; the historical meteorological environment dataset includes historical data subsets for each type of meteorological parameter; the meteorological parameters include temperature, specific humidity, relative humidity, east-west wind speed, north-south wind speed, vertical velocity, geopotential height, 10-meter wind field, 2-meter dew point temperature, 2-meter temperature, sea level pressure, snowfall, surface pressure, total cloud cover, total precipitable water, total water vapor content, and precipitation.
[0007] S2, using the historical meteorological environment dataset, a meteorological training dataset is constructed;
[0008] S3. Based on the meteorological training dataset, the preset cloud and rain field model is processed to obtain a cloud and rain field fusion prediction model.
[0009] S4. Using the cloud and rain field fusion prediction model, the collected meteorological dataset is processed to obtain a set of cloud and rain prediction results.
[0010] The meteorological training dataset constructed using the historical meteorological environment dataset includes:
[0011] S21, preprocess the historical meteorological environment dataset to obtain a preprocessed dataset;
[0012] S22, perform meteorological parameter calculation processing on the preprocessed dataset to obtain a meteorological variable parameter information set; the meteorological variable parameter information set includes a data subset of each meteorological variable; the meteorological variables include the divergence, vorticity, CAPE index, CIN index, K index, and 0-degree layer height of each pressure layer;
[0013] S23, using the snowfall, precipitation, and total cloud cover as label data, and the meteorological variables of all types of meteorological parameters and meteorological variable parameter information sets other than the label data in the preprocessed dataset as training data, perform relevant calculations on the label data and training data to obtain the associated data set of each type of label data;
[0014] S24. Using each type of label data and the corresponding associated data set, a meteorological training dataset is constructed.
[0015] The preprocessing of the historical meteorological and environmental dataset to obtain a preprocessed dataset includes:
[0016] S211, outlier removal processing is performed on the historical meteorological environment dataset to obtain the first dataset;
[0017] S212, Perform time matching processing on the first dataset to obtain the second dataset;
[0018] S213, perform pattern matching processing on the second dataset to obtain a preprocessed dataset.
[0019] The process of performing relevant calculations on the label data and training data to obtain the associated data set for each type of label data includes:
[0020] S231, using the data subsets corresponding to all training data, a first data matrix is constructed; the row vectors of the first data matrix are the data sequences corresponding to the data subsets of each class of training data;
[0021] S232, Estimate the feature dimension of the first data matrix to obtain the feature dimension N;
[0022] S233, based on the feature dimension, perform correlation estimation processing on each type of label data and the training data to obtain the associated data set of the label data.
[0023] The step of estimating the feature dimension N of the first data matrix includes:
[0024] S2321, normalize each row vector of the first data matrix to obtain a first normalized matrix; the row vectors of the first normalized matrix are obtained by normalizing each row vector of the first data matrix.
[0025] S2322, eigenvalues are calculated on the first normalized matrix to obtain an eigenvalue set;
[0026] S2323, Based on the eigenvalue set, perform angle transformation on each column vector of the first normalized matrix to obtain the first transformation sequence;
[0027] S2324, Perform rotation and combination calculations on the first transformation sequence to obtain the first feature matrix;
[0028] S2325, calculate the rank and trace of the first feature matrix;
[0029] S2326, Perform integer eigenvalue calculation on the first feature matrix, rank value and trace value to obtain the first integer value N1;
[0030] S2327, determine the greatest common divisor of the total number of categories of the training data and the first integer value, which is the feature dimension N.
[0031] The expression for the angle transformation is:
[0032] θ i =arccos(x i / γ i ),
[0033] Where, x i γ is the median value of the i-th column vector of the first normalized matrix. i θ represents the eigenvalue corresponding to the i-th column vector of the first normalized matrix. i This is the i-th term of the first transformation sequence;
[0034] The element G in the i-th row and j-th column of the first feature matrix ij The calculation expression is:
[0035] G ij =cos(θ) i +θ j )+sin(θ i -θj ),
[0036] Where, θ j This is the j-th term of the first transformation sequence;
[0037] The expression for calculating the integer eigenvalue is:
[0038]
[0039] In the formula, γmax is the maximum value of the eigenvalues corresponding to all column vectors of the first normalized matrix, ρ and δ are the rank and trace of the first eigenma matrix, respectively, and M is the number of rows of the first eigenma matrix. This indicates rounding down to the nearest integer.
[0040] The process of performing correlation estimation between each type of label data and the training data yields a set of associated data for that type of label data, including:
[0041] S2331, For a subset of labeled data, perform cross-correlation calculation with a subset of training data for each class to obtain the cross-correlation sequence of the training data for each class;
[0042] S2332, using all the cross-correlation sequences as row vectors, construct the first cross-correlation matrix; the index of each row vector corresponds to the index of a class of training data;
[0043] S2333, Perform singular value decomposition on the first cross-correlation matrix to obtain the singular value vector and the corresponding singular values;
[0044] S2334, calculate the correlation between each singular value vector and its corresponding singular value to obtain the corresponding correlation value;
[0045] S2335, Sort all correlation values from high to low, and determine the index of the singular value vector corresponding to the top N correlation values, which is the correlation index set;
[0046] S2336, determine the training data corresponding to all the serial numbers in the relevant serial number set, which are the relevant training data;
[0047] S2337, using a subset of all relevant training data, construct the associated data set of the label data.
[0048] A second aspect of this invention discloses a cloud and rain environment modeling and prediction device based on a supervised machine learning method, the device comprising:
[0049] Memory containing executable program code;
[0050] A processor coupled to the memory;
[0051] The processor calls the executable program code stored in the memory to execute the cloud and rain environment modeling and prediction method based on supervised machine learning.
[0052] In a third aspect of this invention, a computer-storable medium is disclosed, wherein the computer-storable medium stores computer instructions, and when the computer instructions are invoked by a computer, they are used to execute the cloud and rain environment modeling and prediction method based on supervised machine learning.
[0053] In a fourth aspect of this invention, an information data processing terminal is disclosed, which is used to implement the cloud and rain environment modeling and prediction method based on supervised machine learning.
[0054] The beneficial effects of this invention are as follows:
[0055] This invention comprehensively utilizes historical product data and fully considers the influence of various meteorological elements and combined variables on the cloud and rain environment. Through a large amount of training data, it fully approximates the complex nonlinear relationship between key physical parameters of clouds and rain and various elements. The multi-angle and comprehensive consideration of factors enables this model to be applicable to a variety of different situations and has good universality. It can solve the problem of accuracy deviation in generating refined cloud and rain environment data by making full use of various observation data. Attached Figure Description
[0056] Figure 1 This is a flowchart illustrating the implementation of the method of the present invention. Detailed Implementation
[0057] To better understand the content of this invention, an embodiment is provided here.
[0058] Figure 1 This is a flowchart illustrating the implementation of the method of the present invention.
[0059] In a first aspect, this invention discloses a method for modeling and predicting cloud and rain environments based on supervised machine learning, comprising:
[0060] S1, Collect historical meteorological environment dataset; the historical meteorological environment dataset includes historical data subsets for each type of meteorological parameter; the meteorological parameters include temperature, specific humidity, relative humidity, east-west wind speed, north-south wind speed, vertical velocity, geopotential height, 10-meter wind field, 2-meter dew point temperature, 2-meter temperature, sea level pressure, snowfall, surface pressure, total cloud cover, total precipitable water, total water vapor content, and precipitation.
[0061] S2, using the historical meteorological environment dataset, a meteorological training dataset is constructed;
[0062] S3. Based on the meteorological training dataset, the preset cloud and rain field model is processed to obtain a cloud and rain field fusion prediction model.
[0063] S4. Using the cloud and rain field fusion prediction model, the collected meteorological dataset is processed to obtain a set of cloud and rain prediction results.
[0064] The collected meteorological dataset includes temperature, specific humidity, relative humidity, east-west wind speed, north-south wind speed, vertical velocity, geopotential height, 10-meter wind field, 2-meter dew point temperature, 2-meter temperature, sea level pressure, surface pressure, total precipitable water and total water vapor content of each pressure layer, as well as the divergence, vorticity, CAPE index, CIN index, K index, and 0-degree layer height of each pressure layer calculated from these data.
[0065] The height of the pressure layer is the height of a pressure layer with a value of 200, 300, 400, 500, 600, 700, 850, 925, or 1000 hPa.
[0066] The cloud and rain forecast result information set includes the predicted values of snowfall, precipitation, and total cloud cover.
[0067] The meteorological training dataset constructed using the historical meteorological environment dataset includes:
[0068] S21, preprocess the historical meteorological environment dataset to obtain a preprocessed dataset;
[0069] S22, perform meteorological parameter calculation processing on the preprocessed dataset to obtain a meteorological variable parameter information set; the meteorological variable parameter information set includes a data subset of each meteorological variable; the meteorological variables include the divergence, vorticity, CAPE index, CIN index, K index, and 0-degree layer height of each pressure layer;
[0070] The CAPE index, or Convective Available Potential Energy index, is calculated from the temperature and humidity profile (temperature T and dew point distribution with altitude) of a radiosonde station or the thermodynamic field output by a numerical model.
[0071] The CIN index, or Convection Inhibition Energy, represents the energy required to overcome negative buoyancy when an air parcel rises from the ground to the free convection height (LFC). It reflects the difficulty of initiating convection and is calculated using the same method as the CAPE index.
[0072] The K-index is a comprehensive indicator used to assess atmospheric stability and convection potential. It combines temperature, humidity, and vertical gradient, and is calculated using the 850 hPa and 700 hPa layer temperatures (T) and dew point temperatures output from radiosonde stations or models. The 0°C layer height is the altitude at which the atmospheric temperature is 0°C.
[0073] S23, using the snowfall, precipitation, and total cloud cover as label data, and using the historical data subsets of all other types of meteorological parameters and the meteorological variable parameter information set in the preprocessed dataset (excluding label data) as training data, perform relevant calculations on the label data and training data to obtain the associated data set for each type of label data;
[0074] S24. Using each type of label data and the corresponding associated data set, a meteorological training dataset is constructed.
[0075] The data subset is a data sequence of a certain type of data; the data sequence is a collection of data obtained at several acquisition times.
[0076] The preprocessing of the historical meteorological and environmental dataset to obtain a preprocessed dataset includes:
[0077] S211, outlier removal processing is performed on the historical meteorological environment dataset to obtain the first dataset;
[0078] S212, Perform time matching processing on the first dataset to obtain the second dataset;
[0079] S213, perform pattern matching processing on the second dataset to obtain a preprocessed dataset.
[0080] The process of performing relevant calculations on the label data and training data to obtain the associated data set for each type of label data includes:
[0081] S231, using the data subsets corresponding to all training data, a first data matrix is constructed; the row vectors of the first data matrix are the data sequences corresponding to the data subsets of each class of training data;
[0082] S232, Estimate the feature dimension of the first data matrix to obtain the feature dimension N;
[0083] S233, based on the feature dimension, perform correlation estimation processing on each type of label data and the training data to obtain the associated data set of the label data.
[0084] The step of estimating the feature dimension N of the first data matrix includes:
[0085] S2321, normalize each row vector of the first data matrix to obtain a first normalized matrix; the row vectors of the first normalized matrix are obtained by normalizing each row vector of the first data matrix.
[0086] S2322, eigenvalues are calculated on the first normalized matrix to obtain an eigenvalue set;
[0087] S2323, Based on the eigenvalue set, perform angle transformation on each column vector of the first normalized matrix to obtain the first transformation sequence;
[0088] S2324, Perform rotation and combination calculations on the first transformation sequence to obtain the first feature matrix;
[0089] S2325, calculate the rank and trace of the first feature matrix;
[0090] S2326, Perform integer eigenvalue calculation on the first feature matrix, rank value and trace value to obtain the first integer value N1;
[0091] S2327, determine the greatest common divisor of the total number of categories of the training data and the first integer value, which is the feature dimension N.
[0092] The total number of categories in the training data is the total number of data categories in the preprocessed dataset, excluding the label data, of all other types of meteorological parameters and meteorological variable parameter information sets, and its value is 28.
[0093] The expression for the angle transformation is:
[0094] θ i =arccos(x i / γ i ),
[0095] Where, x i γ is the median value of the i-th column vector of the first normalized matrix. i θ represents the eigenvalue corresponding to the i-th column vector of the first normalized matrix. i This is the i-th term of the first transformation sequence;
[0096] The angle transformation, which maps the ratio of the median value to the eigenvalue of the column vector to an angle using an inverse cosine function, can convert the linear relationship of the original meteorological data into an angular relationship in polar coordinates. This facilitates the capture of nonlinear correlations between data (such as the coupling effect between the temperature and humidity gradient of the pressure layer and wind speed). Eigenvalue γ i Reflecting data variance, x i / γ i Normalization can eliminate calculation biases caused by dimensional differences in different meteorological parameters (such as temperature and wind speed), enhancing the comparability of cross-parameter features. The median value is not sensitive to outliers and, compared to the mean, better characterizes the core trend of data distribution. Combined with eigenvalue weighting, it can filter out random noise (such as instrument observation errors) in radiosonde data and retain dominant meteorological features (such as the vertical temperature and humidity gradient of strong convective systems).
[0097] The element G in the i-th row and j-th column of the first feature matrix ij The calculation expression is:
[0098] G ij =cos(θ) i +θ j )+sin(θ i -θ j ),
[0099] Where, θ i and θ j These are the i-th and j-th terms of the first transformation sequence, respectively;
[0100] The first feature matrix uses the sum and difference formulas of trigonometric functions to determine the angle θ between the two features. i θj is converted to the "phase coupling" form, cos(θ) i +θ j Characterizes the synergy between features (e.g., the synchronous change of 850hPa divergence and 500hPa vorticity); sin(θ) i -θ j This expression characterizes the differences between features (such as the asynchronous changes in the vertical temperature lapse rate and CAPE). It can simultaneously capture both linear correlations (such as positive / negative correlations) and nonlinear phase relationships (such as lead / lag effects) of meteorological parameters, making it more suitable for describing the complex coupling of atmospheric dynamic processes than traditional correlation coefficient matrices. Through matrix operations involving angle transformation, the correlations of high-dimensional meteorological parameters (such as 28 types of training data) are compressed into a low-dimensional space, avoiding model overfitting caused by the curse of dimensionality, while preserving the interactions of key physical processes (such as the coordinated changes in temperature, wind field, and humidity during frontal activity).
[0101] The expression for calculating the integer eigenvalue is:
[0102]
[0103] In the formula, γmax is the maximum value of the eigenvalues corresponding to all column vectors of the first normalized matrix, ρ and δ are the rank and trace of the first eigenma matrix, respectively, and M is the number of rows of the first eigenma matrix. This indicates rounding down to the nearest integer.
[0104] Based on the matrix's rank ρ (reflecting feature independence) and trace δ (reflecting the sum of feature variances), and by dynamically adjusting the dimension N1 using the maximum eigenvalue γmax, it can automatically adapt to different meteorological scenarios (such as the feature complexity of different weather processes like heavy rain and snowfall): In severe convective weather, γmax is large, and the expression uses (-γmax) exponentially to suppress high-dimensional noise and focus on core dynamic features (such as CAPE and vertical velocity); in stable weather, the eigenvalue distribution is relatively uniform, and the expression uses the arcsin function for smoothing, avoiding information loss due to excessively low dimensionality. The floor operation combined with the boundedness of trigonometric functions can constrain the dimension N1 to a reasonable range (e.g., 10-20 dimensions), reducing manual parameter tuning costs compared to traditional PCA (which requires manually setting the principal component ratio), while thermodynamic constraints (such as the correlation between γmax and atmospheric stability) ensure the physical meaning of the reduced dimension.
[0105] The process of performing correlation estimation between each type of label data and the training data yields a set of associated data for that type of label data, including:
[0106] For a subset of labeled data, cross-correlation is calculated with a subset of training data for each class to obtain the cross-correlation sequence of the training data for that class.
[0107] The first cross-correlation matrix is constructed by using all the cross-correlation sequences as row vectors; the index of each row vector corresponds to the index of a class of training data.
[0108] Singular value decomposition is performed on the first cross-correlation matrix to obtain the singular value vector and the corresponding singular values;
[0109] For each singular value vector and its corresponding singular value, the correlation is calculated to obtain the corresponding correlation value;
[0110] Sort all relevance values from highest to lowest, and determine the indices of the singular value vectors corresponding to the top N relevance values, which form the relevance index set;
[0111] Determine the training data corresponding to all the indices in the relevant indices set, which are the relevant training data;
[0112] By utilizing a subset of all relevant training data, a related dataset of the labeled data is constructed.
[0113] The correlation between the singular value vector and the corresponding singular value is calculated using the following expression:
[0114]
[0115] Among them, e i γ is the correlation value of the i-th singular value vector. ijμ is the j-th element of the i-th singular value vector. i Let γ be the singular value corresponding to the i-th singular value vector. i M1 is the mean of the i-th singular value vector, and M2 is the number of elements in the singular value vector.
[0116] The correlation calculation between the singular value vector and its corresponding singular value amplifies the weight of elements deviating from the mean in the singular value vector through an exponential function. This highlights features that significantly affect the labeled data (such as snowfall) (e.g., 0°C layer height, total precipitable water volume) and suppresses interference from irrelevant features (such as upper-level wind speed). The denominator utilizes the periodicity of trigonometric functions to map the ratio of the feature mean μi to the vector mean γi as a non-linear weight. When γi / μi approaches π / 2, the denominator approaches 1, at which point the weight of high-deviation elements reaches its maximum, corresponding to the identification of abrupt changes in severe convective weather (such as a sudden increase in CAPE).
[0117] The step of determining the training data corresponding to all the indices in the relevant indices set is to determine the indices of the corresponding row vectors based on all the indices in the relevant indices set, and then determine the indices of the training data based on the indices of the row vectors, thereby determining the training data.
[0118] The cross-correlation sequence is obtained by cross-correlation calculation between the data sequence of a subset of the labeled data and the data sequence of a subset of the training data.
[0119] The cloud and rain field model includes: an input module, a first convolutional module, a second convolutional module, a third convolutional module, a fourth convolutional module, a first normalization module, a first activation module, a second normalization module, a second activation module, a channel splitting module, a channel shuffling module, a first pooling module, a fusion module, and a feature extraction network. The input module receives N input variables. The output of the input module is connected to the input of the first convolutional module, and the output of the first convolutional module is connected to the input of the first pooling module. The output of the first pooling module is connected to the input of the channel splitting module. The output of the channel splitting module is connected to the inputs of the second convolutional module and the fusion module, respectively. The output of the second convolutional module is connected to the input of the first normalization module; the output of the first normalization module is connected to the input of the third convolutional module; the output of the third convolutional module is connected to the input of the first activation module; the output of the first activation module is connected to the input of the fourth convolutional module; the output of the fourth convolutional module is connected to the input of the second normalization module; the output of the second normalization module is connected to the input of the second activation module; the output of the second activation module is connected to the input of the fusion module; the output of the fusion module is connected to the input of the channel shuffling module; the output of the channel shuffling module is connected to the input of the feature extraction network; the output of the feature extraction network is configured as the model output of the target feature extraction model.
[0120] It should be noted that the above-mentioned channel shuffle module is built based on channel shuffle to ensure that the input of the subsequent group convolution comes from different groups, so that information can flow between different groups. This embodiment of the invention does not limit this.
[0121] It should be noted that the above-mentioned channel splitting module is built based on Channel Split to divide into two branches with the same number of channels, and this embodiment of the invention does not limit this.
[0122] It should be noted that the first and third convolutional modules mentioned above are 1×1 ordinary convolutions, with a stride of 2. Furthermore, the second and fourth convolutional modules are 3×3 depthwise convolutions, with a stride of 2, used to achieve feature dimensionality reduction and reduce computational load. This embodiment of the invention does not impose limitations on these modules.
[0123] It should be noted that the fusion module is built based on the Concat operation to double the number of channels and increase the network width, but this embodiment of the invention does not limit it.
[0124] It should be noted that the first pooling module mentioned above is a max pooling layer with a step size of 1, and this embodiment of the present invention does not limit it.
[0125] It should be noted that the first normalization module and the second normalization module mentioned above are constructed based on the batch normalization layer, and this embodiment of the invention does not limit them.
[0126] It should be noted that the first activation module and the second activation module mentioned above are constructed based on the ReLU activation function, and this embodiment of the invention does not limit them.
[0127] It should be noted that the above feature extraction network is constructed based on a bidirectional long short-term memory network, and this embodiment of the invention is not limited thereto. Furthermore, the feature extraction network includes two independent LSTM networks: one responsible for processing the forward information of the sequence, and the other for processing the reverse information, enabling it to learn key features from past and future information, thereby more effectively capturing the dependency features in the time series.
[0128] The cloud and rain field model can be the LightGBM model.
[0129] The time matching process unifies different types of data onto the same time base; the time matching process can employ methods such as extrapolation / extrapolation and Lagrange three-point interpolation.
[0130] The pattern matching process includes:
[0131] S2131, For each data attribute of the second dataset, with the data collection information of the data as the independent variable and the data value of the data as the dependent variable, perform autoregressive-moving average modeling to obtain the regression model of the data attribute of the data class respectively;
[0132] S2132, using the regression model, calculate and process the independent variables to obtain regression data values; determine whether the absolute value of the difference between the regression data value and the corresponding dependent variable value is greater than a set first regression discrimination threshold; if it is greater than the first regression discrimination threshold, delete the data from the second dataset; if it is less than or equal to the first regression discrimination threshold, do not process the data.
[0133] S2133, perform fusion processing on all data in the second dataset after executing S2131 and S2132 to obtain a preprocessed dataset;
[0134] The data collection information refers to the collection time information.
[0135] The autoregressive moving average model can be implemented using an AR-MA model.
[0136] The process of processing the preset cloud-rain field model based on the meteorological training dataset to obtain a cloud-rain field fusion prediction model includes:
[0137] S31, for each type of label data, the associated dataset of the type label data is used as a training data subset, and the data subset corresponding to the type label data is used as a label subset to construct a training dataset for the type label data; the training dataset includes a training data subset and a label subset; each training data in the training data subset has a corresponding label value in the label subset;
[0138] S32, using the training dataset of each type of label data, train the preset cloud and rain field model to obtain the corresponding prediction model;
[0139] S33, using the prediction model of all class label data, a cloud and rain field fusion prediction model is constructed;
[0140] The cloud-rain field fusion prediction model is used to process the collected meteorological dataset to obtain a set of cloud-rain prediction results, including:
[0141] Based on the associated data set of each type of label data, the corresponding meteorological data is obtained from the meteorological dataset. The obtained meteorological data is then input into the corresponding prediction model in the cloud-rain field fusion prediction model to obtain the corresponding prediction result.
[0142] By utilizing all the prediction results, a set of cloud and rain prediction results information is constructed;
[0143] The categories of the prediction results include predictions for snowfall, precipitation, and total cloud cover.
[0144] The process of training a pre-defined cloud and rain field model using the training dataset of each type of label data to obtain the corresponding prediction model includes:
[0145] S321, Divide the training dataset of each type of label data evenly to obtain several training data subsets;
[0146] S322, randomly select a training data subset;
[0147] S323, take each training data in the training data set as input data and input it into the preset cloud and rain field model;
[0148] S324, The input data is processed using the cloud and rain field model to obtain the corresponding predicted value;
[0149] S325, perform evaluation information calculation and processing on the predicted values of all training data and the label values corresponding to the input data to obtain an evaluation information set; the evaluation information set includes evaluation information; the evaluation information is accuracy, hit rate, false alarm rate, missed alarm rate, and TS score.
[0150] S326, determine whether the evaluation information set satisfies the convergence condition, and obtain the first judgment result;
[0151] When the first judgment result is negative, the model training state is determined to be that the termination training condition is not met.
[0152] When the first judgment result is yes, it is determined that the model training state meets the termination training condition;
[0153] S327, when the model training state does not meet the termination training condition, the parameters of the cloud and rain field model are updated using the parameter update model, triggering the execution of S322;
[0154] When the model training state meets the termination training condition, the training process of the cloud and rain field model is completed, and the corresponding prediction model is obtained.
[0155] The convergence condition of the evaluation information set means that each evaluation information in the evaluation information set is greater than the corresponding preset discrimination threshold.
[0156] The expression for calculating the precision is:
[0157]
[0158] In the formula, OA represents the accuracy, TP represents the number of training data that were actually correctly detected, and FP represents the number of training data that were actually incorrectly detected.
[0159] Hit rate: In the formula, NA represents the number of correct predictions for positive samples, and NC represents the number of missed predictions.
[0160] Reverse air report rate: In the formula, NB represents the number of empty reports.
[0161] Inverse false negative rate:
[0162] TS rating:
[0163] The preset thresholds for accuracy, hit rate, false alarm rate, missed alarm rate, and TS score can be 0.8, 0.7, 0.6, 0.75, and 0.9, respectively.
[0164] The collected meteorological dataset can be obtained using micro-rain radar, raindrop spectrometer, fogdrop spectrometer, visibility meter, millimeter-wave radar, dual-polarization Doppler radar, wind profiler radar, microwave radiometer, active lightning positioning imaging system, and ground electric field meter.
[0165] The expression for the parameter update model is:
[0166]
[0167] Where, x (i) For the i-th training data in the training data set, y (i) The label value of the i-th training data in the training data subset. Let v be the loss function, θ be the parameter update value, and θ be the parameters of the cloud-rain field model. This represents taking the partial derivative with respect to the variable θ, f(x) (i) ;θ) represents the predicted value obtained by the cloud and rain field model for the i-th training data in the training data set, f(.) is the calculation function corresponding to the cloud and rain field model; exp represents the power operation of the constant e.
[0168] A second aspect of this invention discloses a cloud and rain environment modeling and prediction device based on a supervised machine learning method, the device comprising:
[0169] Memory containing executable program code;
[0170] A processor coupled to the memory;
[0171] The processor calls the executable program code stored in the memory to execute the cloud and rain environment modeling and prediction method based on supervised machine learning.
[0172] In a third aspect of this invention, a computer-storable medium is disclosed, wherein the computer-storable medium stores computer instructions, and when the computer instructions are invoked by a computer, they are used to execute the cloud and rain environment modeling and prediction method based on supervised machine learning.
[0173] In a fourth aspect of this invention, an information data processing terminal is disclosed, which is used to implement the cloud and rain environment modeling and prediction method based on supervised machine learning.
[0174] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.
Claims
1. A cloud and rain environment modeling and prediction method based on supervised machine learning, characterized in that, include: S1, Collect historical meteorological environment dataset; the historical meteorological environment dataset includes historical data subsets for each type of meteorological parameter; the meteorological parameters include temperature, specific humidity, relative humidity, east-west wind speed, north-south wind speed, vertical velocity, geopotential height, 10-meter wind field, 2-meter dew point temperature, 2-meter temperature, sea level pressure, snowfall, surface pressure, total cloud cover, total precipitable water, total water vapor content, and precipitation. S2, using the aforementioned historical meteorological dataset, a meteorological training dataset is constructed, including: S21, preprocess the historical meteorological environment dataset to obtain a preprocessed dataset; S22, perform meteorological parameter calculation processing on the preprocessed dataset to obtain a meteorological variable parameter information set; the meteorological variable parameter information set includes a data subset of each meteorological variable; the meteorological variables include the divergence, vorticity, CAPE index, CIN index, K index, and 0-degree layer height of each pressure layer; S23, using the snowfall, precipitation, and total cloud cover as label data, and the meteorological variables from the set of meteorological parameters and meteorological variable parameter information of all other types of meteorological parameters and meteorological variable parameters in the preprocessed dataset (excluding the label data) as training data, perform relevant calculations on the label data and training data to obtain the associated data set for each type of label data, including: S231, using the data subsets corresponding to all training data, a first data matrix is constructed; the row vectors of the first data matrix are the data sequences corresponding to the data subsets of each class of training data; S232, Estimate the feature dimension of the first data matrix to obtain the feature dimension N; S233, Based on the feature dimension, perform correlation estimation processing on each type of label data and the training data to obtain the associated data set of the label data; S24. Using each type of label data and the corresponding associated data set, a meteorological training dataset is constructed. S3. Based on the meteorological training dataset, the preset cloud and rain field model is processed to obtain a cloud and rain field fusion prediction model. S4. Using the cloud and rain field fusion prediction model, the collected meteorological dataset is processed to obtain a set of cloud and rain prediction results.
2. The cloud and rain environment modeling and prediction method based on supervised machine learning as described in claim 1, characterized in that, The preprocessing of the historical meteorological and environmental dataset to obtain a preprocessed dataset includes: S211, outlier removal processing is performed on the historical meteorological environment dataset to obtain the first dataset; S212, Perform time matching processing on the first dataset to obtain the second dataset; S213, perform pattern matching processing on the second dataset to obtain a preprocessed dataset.
3. The cloud and rain environment modeling and prediction method based on supervised machine learning as described in claim 1, characterized in that, The step of estimating the feature dimension N of the first data matrix includes: S2321, normalize each row vector of the first data matrix to obtain a first normalized matrix; the row vectors of the first normalized matrix are obtained by normalizing each row vector of the first data matrix. S2322, eigenvalues are calculated on the first normalized matrix to obtain an eigenvalue set; S2323, Based on the eigenvalue set, perform angle transformation on each column vector of the first normalized matrix to obtain the first transformation sequence; S2324, Perform rotation and combination calculations on the first transformation sequence to obtain the first feature matrix; S2325, calculate the rank and trace of the first feature matrix; S2326, Perform integer eigenvalue calculation on the first feature matrix, rank, and trace to obtain the first integer value. ; S2327, determine the greatest common divisor of the total number of categories of the training data and the first integer value, which is the feature dimension N.
4. The cloud and rain environment modeling and prediction method based on supervised machine learning as described in claim 3, characterized in that, The expression for the angle transformation is: ， in, Let be the median value of the i-th column vector of the first normalized matrix. The eigenvalues are the eigenvalues corresponding to the i-th column vector of the first normalized matrix. This is the i-th term of the first transformation sequence; The elements in the i-th row and j-th column of the first feature matrix The calculation expression is: , in, This is the j-th term of the first transformation sequence; The expression for calculating the integer eigenvalue is: ， In the formula, The maximum value of the eigenvalues corresponding to all column vectors of the first normalized matrix. and Let M be the rank and trace of the first feature matrix, respectively, and M be the number of rows in the first feature matrix. This indicates rounding down to the nearest integer.
5. The cloud and rain environment modeling and prediction method based on supervised machine learning as described in claim 3, characterized in that, The process of performing correlation estimation between each type of label data and the training data yields a set of associated data for that type of label data, including: S2331, For a subset of labeled data, perform cross-correlation calculation with a subset of training data for each class to obtain the cross-correlation sequence of the training data for each class; S2332, using all the cross-correlation sequences as row vectors, construct the first cross-correlation matrix; the index of each row vector corresponds to the index of a class of training data; S2333, Perform singular value decomposition on the first cross-correlation matrix to obtain the singular value vector and the corresponding singular values; S2334, calculate the correlation between each singular value vector and its corresponding singular value to obtain the corresponding correlation value; S2335, Sort all correlation values from high to low, and determine the index of the singular value vector corresponding to the top N correlation values, which is the correlation index set; S2336, determine the training data corresponding to all the serial numbers in the relevant serial number set, which are the relevant training data; S2337, using a subset of all relevant training data, construct the associated data set of the label data.
6. A cloud and rain environment modeling and prediction device based on supervised machine learning methods, characterized in that, The device includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the cloud and rain environment modeling and prediction method based on supervised machine learning as described in any one of claims 1 to 5.
7. A computer-storable medium, characterized in that, The computer storage medium stores computer instructions, which, when invoked by the computer, are used to execute the cloud and rain environment modeling and prediction method based on supervised machine learning as described in any one of claims 1 to 5.
8. An information data processing terminal, characterized in that, The information data processing terminal is used to implement the cloud and rain environment modeling and prediction method based on supervised machine learning as described in any one of claims 1 to 5.