Short-time rainfall prediction method based on RandomForest and LightGBM algorithms

Abstract

The invention discloses a short-time rainfall prediction method based on RandomForest and LightGBM algorithms. The method comprises the following steps: acquiring rainfall observation data and grid data predicted by a numerical weather mode; interpolating the grid data predicted by the numerical weather mode to a meteorological automatic station according to a neighbor method; the method comprises the following steps: preprocessing rainfall observation live data, geographic information and grid data predicted by a numerical weather mode of an automatic meteorological station, and establishing a prediction data set in a table form; calculating a Spearman correlation coefficient, performing feature selection by using an information gain method, and constructing a prediction data set; according to the prediction data set, using a LightGBM algorithm framework to carry out regression prediction model training, and using a RandomForest algorithm framework to carry out classification prediction model training; and performing evaluation and parameter optimization on the prediction model. According to the method, the prediction model is adjusted according to recent rainfall climate change characteristics, and prediction deviation caused by climate change is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of precipitation prediction, and more specifically to a short-term precipitation prediction method based on RandomForest and LightGBM algorithms. Background Art

[0002] Precipitation forecast is one of the most important meteorological forecast services, which includes the prediction of precipitation amount and precipitation area. The numerical weather forecast data currently used is the prediction result for the future IT+1H to IT+36H, but due to the network delay of data transmission, only the numerical results for IT+12H to IT+36H can be used. Due to the influence of factors such as the initial field deviation and different parameterization schemes of the numerical weather forecast model, there is a certain deviation between the precipitation forecast and the observed value. Summary of the invention

[0003] To this end, the technical problem to be solved by the present invention is to provide a short-term precipitation prediction method based on RandomForest and LightGBM algorithms, which can obtain hourly precipitation prediction within the next 24 hours.

[0004] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0005] A short-term precipitation prediction method based on RandomForest and LightGBM algorithm includes the following steps:

[0006] P1. Obtain hourly precipitation observation data from automatic meteorological stations and grid data predicted by numerical weather models;

[0007] P2, interpolate the grid data predicted by the numerical weather model to the meteorological automatic station according to the nearest neighbor method;

[0008] P3, pre-processing the actual precipitation observation data of the meteorological automatic station, geographic information and grid data predicted by the numerical weather model, and constructing a prediction data set in a tabular form according to the selected training date;

[0009] P4. Calculate the Spearman correlation coefficients of various elements in the numerical weather model prediction grid data and the automatic station precipitation observation data, and use the information gain method to select features to construct the prediction data set;

[0010] P5. Based on the prediction data set, the LightGBM algorithm framework is used to train the regression prediction model, and the RandomForest algorithm framework is used to train the classification prediction model;

[0011] P6. Use the validation data set to evaluate the prediction model and optimize the parameters, and finally obtain and store the precipitation prediction model.

[0012] In the above-mentioned short-term precipitation prediction method based on RandomForest and LightGBM algorithms, in step P1, the study area of ​​the hourly meteorological automatic station precipitation observation data and the grid data predicted by the numerical weather model is 70.95°E~100.0°E, 31.80°N~51.65°N;

[0013] In step P2, the nearest neighbor method is used to extract the forecast data of various meteorological elements of the grid points closest to the automatic precipitation observation site in the numerical weather model grid data, which is used to construct the forecast training set.

[0014] In the above-mentioned short-term precipitation prediction method based on RandomForest and LightGBM algorithms, in step P3, the actual precipitation observation data of the meteorological automatic station, geographic information and numerical weather prediction model data are preprocessed to eliminate or correct erroneous data (the grid data predicted by the numerical weather model sometimes have negative precipitation values, which should be corrected to 0. When the precipitation data is greater than a certain threshold, the value is reduced to a reasonable range or directly eliminated, or the missing data of a certain time is eliminated), and the ID, reporting time IT and prediction time FT of the meteorological automatic station are matched, and then the numerical weather model prediction data of the past 30 days are selected to construct a prediction data set; and the prediction data set is randomly divided into two parts, of which 75% is used as a training data set and 25% is used as a verification data set.

[0015] In the above-mentioned short-term precipitation prediction method based on RandomForest and LightGBM algorithms, the actual precipitation observation data of the meteorological automatic station is the precipitation observation data; the geographic information is the terrain parameters of the meteorological automatic station, including longitude, latitude and altitude difference parameters between the numerical weather prediction model and the station; the prediction elements of the numerical weather prediction model data are the atmospheric temperature, relative humidity, specific humidity, horizontal wind speed U and V components of each pressure layer between 500-925hPa at the FT moment, as well as the vertical wind speed, radar reflectivity and precipitation at the FT moment, and the precipitation at the FT-1H moment.

[0016] The above short-term precipitation prediction method based on RandomForest and LightGBM algorithm, in P4, includes:

[0017] P4-1. Calculate the Spearman correlation coefficients between various elements in the numerical weather model prediction grid data and the automatic station precipitation observation data, and obtain the absolute value of the Spearman correlation coefficient ≥ 0.2, and the number of related element features with the highest absolute value of the Spearman correlation coefficient does not exceed 5;

[0018] P4-2. Based on the information gain screening method, the five regression element features of the regression prediction model and the five classification element features of the classification prediction model were screened for various elements in the numerical weather model prediction grid data and the automatic station precipitation observation data;

[0019] P4-3. The correlation factor feature with the highest Spearman correlation coefficient obtained in step P4-1 is first combined with the five regression factor features of the regression prediction model obtained in step P4-2 to obtain a regression feature data set; and then combined with the five classification factor features of the classification prediction model to obtain a classification feature data set.

[0020] In the above-mentioned short-term precipitation prediction method based on RandomForest and LightGBM algorithms, in step P4-1, when calculating the related element feature with the highest absolute value of the Spearman correlation coefficient, it is necessary to satisfy that the correlation coefficient between the related element features with the highest absolute value of the Spearman correlation coefficient is ≤0.6. If the above condition is not met, the feature is eliminated and the next feature is searched.

[0021] In the above short-term precipitation prediction method based on RandomForest and LightGBM algorithm, in step P5, according to the modeling prediction feature data set,

[0022] P5-1. For the regression feature data set, use the LightGBM algorithm framework to build the LightGBM algorithm regression prediction model, and train the LightGBM algorithm regression prediction model to obtain the regression prediction result R i ;

[0023] P5-2. For the classification feature data set, use the RandomForest algorithm framework to build a classification prediction model, and train the classification prediction model to obtain the classification prediction result C i ; No precipitation C i =0, when there is precipitation C i =1;

[0024] P5-3. Regression prediction result R i And the classification prediction result C i The integration result is the integrated prediction result E i ;

[0025] P5-4. Obtain the regression feature data of the regression prediction model and the classification feature data of the classification prediction model for the future IT+12H to IT+36H from the numerical weather model prediction grid data, bring the regression feature data into the LightGBM algorithm regression prediction model, and obtain the regression prediction result R i; Bring the classification feature data into the RandomForest algorithm classification prediction model to obtain the classification prediction result C i ; Then the regression prediction result R i And the classification prediction result C i Perform integration to obtain the integrated result E i , output the short-term precipitation forecast results for the future IT+1H to IT+24H.

[0026] The above short-term precipitation prediction method based on RandomForest and LightGBM algorithm has the following integration rules:

[0027] When C i =1, if R i >=0.1 then E i =R i , if R i <0.1 then E i =0.1;

[0028] When C i = 0, if R i <2.0, E i =0, if R i >=2.0,E i =0.1;

[0029] Among them, i represents a certain prediction time in the future.

[0030] In the above short-term precipitation prediction method based on RandomForest and LightGBM algorithms, in step P6, the RandomForest algorithm classification prediction model adopts the PC scoring standard:

[0031] PC=(NA+ND) / (NA+NB+NC+ND) (1),

[0032] In formula (1), NA is the number of correct precipitation predictions, NB is the number of false precipitation reports, NC is the number of missed precipitation reports, and ND is the number of correct no precipitation predictions.

[0033] The above short-term precipitation prediction method based on RandomForest and LightGBM algorithms, the LightGBM algorithm regression prediction model adopts the MAE scoring standard,

[0034]

[0035] Among them, F is the predicted value, O is the observed value, i is a certain prediction time, and n is the sum of all prediction times.

[0036] The technical solution of the present invention achieves the following beneficial technical effects:

[0037] 1. The present invention uses RandomForest and LightGBM models to construct a short-term precipitation prediction model, and uses the prediction model to realize the hourly precipitation prediction of the observation station in the next 24 hours. The normal operation time of the model can be controlled within 15 minutes. The method of the present invention can adjust the prediction model according to the recent precipitation climate characteristics, reduce the prediction deviation caused by climate change, significantly reduce the calculation speed and operation and maintenance costs, and effectively improve the prediction accuracy and prediction timeliness.

[0038] This scheme uses numerical weather forecast data and is based on the RandomForest and LightGBM algorithms to improve the prediction accuracy of precipitation amount and precipitation area, which has better guiding significance for short-term precipitation forecasting.

[0039] 2. This application constructs a short-term precipitation prediction model using a machine learning method, which includes two steps: classification and regression. The classification model uses the "RandomForest" classification prediction method, and the regression model uses the "LightGBM" regression prediction method.

[0040] 3. The relevant element features obtained by the Spearman correlation coefficient in this application are respectively combined with the 5 regression element features of the regression prediction model and the 5 classification element features of the classification prediction model to obtain a regression feature data set and a classification feature data set, respectively. The Spearman correlation coefficient can be used to supplement the screened regression element features and classification element features, so that the obtained regression feature data set and classification feature data set can better reflect the precipitation prediction requirements.

[0041] 4. The short-term precipitation prediction method of the present application is designed based on the precipitation prediction scoring method. The evaluation indicators of precipitation prediction include: PC, MAE, RMSE and other evaluation indicators. Among them, the PC indicator is to evaluate the proportion of the number of correct precipitation predictions to the total number of predictions. Therefore, when the evaluation function is the PC indicator, the prediction accuracy belongs to a classification problem, which reflects the accuracy of the precipitation area. The MAE and RMSE indicators reflect the error between the observed value (true value) and the predicted value. Since RMSE will excessively amplify the prediction error of heavy precipitation, the MAE indicator is selected in this prediction method to evaluate the error between the observed value and the predicted value, which reflects the accuracy of the precipitation.

[0042] 5. The RandomForest and LightGBM prediction methods have the following advantages: high accuracy; insensitive to extreme data; fast operation speed; no need for data scaling or normalization; prevent overfitting. The model has the disadvantages of relying on high-performance numerical prediction models; in seasons with less precipitation, the proportion of samples with and without precipitation will be seriously unbalanced, and the accuracy of RandomForest classification prediction will be reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A flow chart of a short-term precipitation prediction method based on RandomForest and LightGBM algorithms of the present invention;

[0044] Figure 2 Map of some meteorological stations in Xinjiang;

[0045] Figure 3a This is the PC evaluation distribution map of the short-term precipitation prediction method of this application; Figure 3b It is the PC evaluation distribution diagram of the numerical model CMA;

[0046] Figure 4a This is the MAE evaluation distribution map of the short-term precipitation prediction method in this application; Figure 4b It is the MAE evaluation distribution diagram of the numerical model CMA;

[0047] Figure 5 Comparison chart of the prediction results of the short-term precipitation prediction method in this application for 1H to 24H and the numerical model CMA prediction results for the same prediction period. DETAILED DESCRIPTION

[0048] The short-term precipitation prediction method based on RandomForest and LightGBM algorithm in this application specifically includes the following steps, see Figure 1 .

[0049] P1. Obtain hourly precipitation observation data from automatic meteorological stations and grid data predicted by numerical weather models;

[0050] Xinjiang has different topography and landforms, and many factors need to be considered, including seasonal changes, topography, climate change trends, etc. The research area of ​​the hourly meteorological automatic station precipitation observation data and numerical weather model forecast grid data in this application is 70.95°E~100.0°E, 31.80°N~51.65°N. Figure 2 These are some meteorological stations in Xinjiang.

[0051] Numerical weather model: This application uses the China Meteorological Administration regional mesoscale numerical model CMA_MESO 3km.

[0052] P2, interpolate the grid data predicted by the numerical weather model to the meteorological automatic station according to the nearest neighbor method;

[0053] The nearest neighbor method is used to extract the forecast data of various meteorological elements closest to the automatic precipitation observation station in the numerical weather model forecast, which is used to construct the forecast training set.

[0054] P3, pre-processing the actual precipitation observation data of the meteorological automatic station, geographic information and grid data predicted by the numerical weather model, and constructing a prediction data set in a tabular form according to the selected training date;

[0055] The actual precipitation observation data of the automatic meteorological station, geographic information and numerical weather forecast model data are preprocessed to eliminate or correct erroneous data (the grid data predicted by the numerical weather model sometimes have negative precipitation values, which should be corrected to 0. When the precipitation data is greater than a certain threshold, the value is reduced to a reasonable range or directly eliminated, or the missing data of a certain time period is eliminated), and matched according to the ID of the automatic meteorological station, the reporting time IT and the forecast time FT, and then the numerical weather model forecast data of the past 30 days are selected to construct the forecast data set; and the forecast data set is randomly divided into two parts, of which 75% is used as the training data set and 25% is used as the verification data set.

[0056] The actual precipitation observation data of the automatic meteorological station is the precipitation observation data; the geographic information is the terrain parameters of the automatic meteorological station, including longitude, latitude and the altitude difference parameters between the numerical weather prediction model and the station; the prediction elements of the numerical weather prediction model data are the atmospheric temperature, relative humidity, specific humidity, horizontal wind speed U and V components of each pressure layer between 500-925hPa at the FT time, as well as the vertical wind speed, radar reflectivity and precipitation at the FT time, and precipitation at the FT-1H time. (This prediction element is the characteristic element that is most relevant to precipitation.)

[0057] P4. Calculate the Spearman correlation coefficients of various elements in the numerical weather model prediction grid data and the automatic station precipitation observation data, and use the information gain method to select features to construct the prediction data set;

[0058] P4-1. Calculate the Spearman correlation coefficient of various elements in the numerical weather model prediction grid data and the automatic station precipitation observation data, and obtain the Spearman correlation coefficient absolute value ≥ 0.2, and the number of related element features with the highest absolute value of the Spearman correlation coefficient does not exceed 5; when calculating the related element feature with the highest absolute value of the Spearman correlation coefficient, it is necessary to satisfy that the correlation coefficient between the related element features with the highest absolute value of the Spearman correlation coefficient is ≤ 0.6. If the above conditions are not met, the feature is eliminated and the search for the next feature continues.

[0059] P4-2. Based on the information gain screening method, the five regression element features of the regression prediction model and the five classification element features of the classification prediction model were screened for various elements in the numerical weather model prediction grid data and the automatic station precipitation observation data;

[0060] P4-3. The correlation factor feature with the highest Spearman correlation coefficient obtained in step P4-1 is first combined with the five regression factor features of the regression prediction model obtained in step P4-2 to obtain a regression feature data set; and then combined with the five classification factor features of the classification prediction model to obtain a classification feature data set.

[0061] In this embodiment:

[0062] The characteristics of the correlated factors with the highest absolute values ​​of the Spearman correlation coefficients were: precipitation, radar reflectivity, and 2-meter relative humidity.

[0063] Regression factor characteristics of the regression prediction model: precipitation, 2-meter relative humidity, radar reflectivity, 10-meter U wind, and 10-meter V wind.

[0064] Classification factor characteristics of the classification prediction model: precipitation, radar reflectivity, 2-meter relative humidity, maximum vertical helicity, and 10-meter V wind.

[0065] The Spearman correlation coefficient was unioned with the regression factor features to obtain the regression feature data set: precipitation, radar reflectivity, 2-meter relative humidity, 10-meter U wind, and 10-meter V wind.

[0066] The Spearman correlation coefficient was unioned with the classification element features to obtain the classification feature data set: precipitation, radar reflectivity, 2-meter relative humidity, maximum vertical helicity, and 10-meter V wind.

[0067] P5. Based on the prediction data set, the LightGBM algorithm framework is used to train the regression prediction model, and the RandomForest algorithm framework is used to train the classification prediction model;

[0068] P5-1. For the regression feature data set, use the LightGBM algorithm framework to build the LightGBM algorithm regression prediction model, and train the LightGBM algorithm regression prediction model to obtain the regression prediction result R i ;

[0069] LightGBM is a gradient boosting algorithm based on a tree model. In the regression problem of predicting precipitation, LightGBM will build multiple regression decision trees. Each decision tree is learned by optimizing the loss function (such as mean absolute error MAE or root mean square error RMSE) based on the training data. LightGBM will perform a weighted summation of the prediction results of all decision trees and obtain the final precipitation prediction value.

[0070] P5-2. For the classification feature data set, use the RandomForest algorithm framework to build the RandomForest algorithm classification prediction model, and train the RandomForest algorithm classification prediction model to obtain the classification prediction result C i ;

[0071] RandomForest is an ensemble learning method. It builds multiple decision tree models and then averages or votes on the predictions of these tree models to get the final prediction. For example, this method is used to predict whether it will rain in the future. The predictions of all decision tree models will vote, and if most decision trees predict rain, then RandomForest will also predict that it will rain in the future.

[0072] P5-3, No precipitation C i =0, when there is precipitation C i =1; the regression prediction result R i And the classification prediction result C i The integration result is the integrated prediction model E i , which is the short-term precipitation prediction model;

[0073] When C i =1, if R i >=0.1 then E i =R i , if R i <0.1 then E i =0.1;

[0074] When C i = 0, if R i <2.0, E i =0, if R i >=2.0,E i =0.1;

[0075] Among them, i represents a certain prediction time in the future.

[0076] Integration of classification and regression prediction results. Sometimes the classification prediction and regression prediction results are inconsistent. For example, the classification prediction result is 0 (indicating no precipitation), while the regression prediction result is greater than 0 (indicating precipitation), and the two prediction results are contradictory. Therefore, this step can be used to correct the final prediction result.

[0077] Prediction result R i And the classification prediction result C i E with integrated results i The obtained PC and MAE results are consistent.

[0078] P5-4. Obtain the regression feature data of the regression prediction model and the classification feature data of the classification prediction model for the future IT+12H to IT+36H from the numerical weather model prediction grid data, bring the regression feature data into the LightGBM algorithm regression prediction model, and obtain the regression prediction result R i ; Bring the classification feature data into the RandomForest algorithm classification prediction model to obtain the classification prediction result C i ; Then the regression prediction result R i And the classification prediction result C i Perform integration to obtain the integrated result E i In step P5-3, the integrated algorithm outputs the short-term precipitation forecast results from IT+1H to IT+24H.

[0079] P6. Use the validation data set to evaluate the prediction model and optimize the parameters, and finally obtain and store the precipitation prediction model.

[0080] The RandomForest algorithm classification prediction model uses the PC scoring standard:

[0081] PC=(NA+ND) / (NA+NB+NC+ND) (1),

[0082] In formula (1), NA is the number of correct precipitation predictions, NB is the number of false precipitation reports, NC is the number of missed precipitation reports, and ND is the number of correct no precipitation predictions.

[0083] The LightGBM algorithm regression prediction model uses the MAE scoring standard.

[0084]

[0085] Among them, F is the predicted value, O is the observed value, i is a certain prediction time, and n is the sum of all prediction times.

[0086] like Figure 3a and Figure 3b They are the PC evaluation distribution diagrams of the short-term precipitation prediction method and the numerical model CMA; Figure 4a and Figure 4b These are the MAE evaluation distribution maps of the short-term precipitation prediction method and the numerical model CMA.

[0087] Figure 5 Table 1 shows the 1-hour precipitation prediction scores of the short-term precipitation prediction method and the numerical model CMA.

[0088] Table 1 1-h precipitation prediction scores of model FCT and numerical model CMA

[0089] Prediction Model PC MAE FCT 0.9645 0.041 CMA 0.9583 0.0424

[0090] It can be seen from Table 1 that compared with the prediction results of the regional mesoscale numerical model CMA_MESO 3km of the China Meteorological Administration, the PC and MAE of the short-term precipitation prediction method based on the RandomForest and LightGBM algorithms in this application are better than the CMA numerical prediction model. The PC value is [0,1]. The larger the PC value, the higher the prediction accuracy frequency; the MAE value is [0,+∞]. The smaller the MAE, the smaller the deviation from the observed value.

[0091] This application uses the PC of the RandomForest algorithm classification prediction model to score the accuracy of sunny and rainy forecasts (the larger the PC value, the higher the accuracy of precipitation prediction), and uses the mean absolute error MAE to score the accuracy of the LightGBM algorithm regression prediction model (the smaller the MAE, the closer it is to the true value). That is, in this application, the short-term precipitation prediction model E i That is, the accuracy of precipitation frequency and numerical precision are taken into account.

[0092] This is an effective evaluation method specifically for Xinjiang. Xinjiang is located in an arid area, and its average annual precipitation frequency and precipitation are much lower than other parts of my country. According to statistics in the past five years, the precipitation frequency of most stations in Xinjiang ≥ 0.1 mm is less than 5% (that is, precipitation occurs in less than 5 hours within 100 hours), and the precipitation frequency in southern Xinjiang is even lower. In addition to considering the accuracy MAC, the accuracy of sunny and rainy forecasts PC is a key prediction evaluation indicator in Xinjiang. The short-term precipitation prediction method based on RandomForest and LightGBM algorithms in this application evaluates the short-term precipitation prediction model from two perspectives, namely, considering the accuracy of precipitation frequency and numerical accuracy.

[0093] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the claims of this patent application.

Claims

1. A short-term precipitation prediction method based on RandomForest and LightGBM algorithm, characterized in that: The steps include: P1. Obtain hourly precipitation observation data from automatic meteorological stations and grid data predicted by numerical weather models; P2, interpolate the grid data predicted by the numerical weather model to the meteorological automatic station according to the nearest neighbor method; P3, pre-processing the actual precipitation observation data of the meteorological automatic station, geographic information and grid data predicted by the numerical weather model, and constructing a prediction data set in a tabular form according to the selected training date; P4. Calculate the Spearman correlation coefficients of various elements in the numerical weather model prediction grid data and the automatic station precipitation observation data, and use the information gain method to select features to construct the prediction data set; P5. Based on the prediction data set, the LightGBM algorithm framework is used to train the regression prediction model, and the RandomForest algorithm framework is used to train the classification prediction model; P6. Use the validation data set to evaluate the prediction model and optimize the parameters, and finally obtain and store the precipitation prediction model.

2. The short-term precipitation prediction method based on RandomForest and LightGBM algorithm according to claim 1, characterized in that: In step P1, the study area of ​​hourly meteorological automatic station precipitation observation data and grid data predicted by numerical weather model is 70.95°E~100.0°E, 31.80°N~51.65°N; In step P2, the nearest neighbor method is used to extract the forecast data of various meteorological elements closest to the grid point of the automatic precipitation observation station in the numerical weather model forecast, which is used to construct the forecast training set.

3. The short-term precipitation prediction method based on RandomForest and LightGBM algorithm according to claim 1, characterized in that: In step P3, the actual precipitation observation data of the automatic meteorological station, the geographic information and the numerical weather forecast model data are preprocessed to eliminate or correct the erroneous data, and matched according to the ID of the automatic meteorological station site, the reporting time IT and the forecast time FT, and then the numerical weather model forecast data of the past 30 days are selected to construct the forecast data set; and the forecast data set is randomly divided into two parts, of which 75% is used as the training data set and 25% is used as the verification data set.

4. The short-term precipitation prediction method based on RandomForest and LightGBM algorithm according to claim 3 is characterized in that: The actual precipitation observation data of the automatic meteorological station are the precipitation observation data; the geographical information are the terrain parameters of the automatic meteorological station, including the longitude, latitude and the altitude difference parameters between the numerical weather prediction model and the station; the prediction elements of the numerical weather prediction model data are the atmospheric temperature, relative humidity, specific humidity, horizontal wind speed U and V components of each pressure layer between 500-925hPa at the FT moment, as well as the vertical wind speed, radar reflectivity and precipitation at the FT moment, and the precipitation at the FT-1H moment.

5. The short-term precipitation prediction method based on RandomForest and LightGBM algorithm according to claim 1 is characterized in that: In P4, it includes: P4-1. Calculate the Spearman correlation coefficients between various elements in the numerical weather model prediction grid data and the automatic station precipitation observation data, and obtain the absolute value of the Spearman correlation coefficient ≥ 0.2, and the number of related element features with the highest absolute value of the Spearman correlation coefficient does not exceed 5; P4-2. Based on the information gain screening method, the five regression element features of the regression prediction model and the five classification element features of the classification prediction model were screened for various elements in the numerical weather model prediction grid data and the automatic station precipitation observation data; P4-3. The correlation factor feature with the highest Spearman correlation coefficient obtained in step P4-1 is first combined with the five regression factor features of the regression prediction model obtained in step P4-2 to obtain a regression feature data set; and then combined with the five classification factor features of the classification prediction model to obtain a classification feature data set.

6. The short-term precipitation prediction method based on RandomForest and LightGBM algorithm according to claim 5 is characterized in that: In step P4-1, when calculating the correlation factor feature with the highest absolute value of the Spearman correlation coefficient, the correlation coefficient between the correlation factor features with the highest absolute value of the Spearman correlation coefficient must be ≤0.

6. If the above condition is not met, the feature is eliminated and the next feature is searched for.

7. The short-term precipitation prediction method based on RandomForest and LightGBM algorithm according to claim 5, characterized in that: In step P5, a feature data set is predicted based on the modeling. P5-1. For the regression feature data set, use the LightGBM algorithm framework to build the LightGBM algorithm regression prediction model, and train the LightGBM algorithm regression prediction model to obtain the regression prediction result R i ; P5-2. For the classification feature data set, use the RandomForest algorithm framework to build a classification prediction model, and train the classification prediction model to obtain the classification prediction result C i ; No precipitation C i =0, when there is precipitation C i =1; P5-3. Regression prediction result R i And the classification prediction result C i The integration result is the integrated prediction result E i ; P5-4. Obtain the regression feature data of the regression prediction model and the classification feature data of the classification prediction model for the future IT+12H to IT+36H from the numerical weather model prediction grid data, bring the regression feature data into the LightGBM algorithm regression prediction model, and obtain the regression prediction result R i ; Bring the classification feature data into the RandomForest algorithm classification prediction model to obtain the classification prediction result C i ; Then the regression prediction result R i And the classification prediction result C i Perform integration to obtain the integrated result E i , output the short-term precipitation forecast results for the next 1H to 24H.

8. The short-term precipitation prediction method based on RandomForest and LightGBM algorithm according to claim 7 is characterized in that: The integration rules are as follows, When C i =1, if R i >=0.1 then E i =R i , if R i <0.1 then E i =0.1; When C i = 0, if R i <2.0, E i =0, if R i >=2.0,E i =0.1; Among them, i represents a certain prediction time in the future.

9. The short-term precipitation prediction method based on RandomForest and LightGBM algorithm according to claim 7, characterized in that: In step P6, the RandomForest algorithm classification prediction model uses the PC scoring standard for the accuracy of sunny and rainy predictions: PC=(NA+ND) / (NA+NB+NC+ND) (1), In formula (1), NA is the number of correct precipitation predictions, NB is the number of false precipitation reports, NC is the number of missed precipitation reports, and ND is the number of correct no precipitation predictions.

10. The short-term precipitation prediction method based on RandomForest and LightGBM algorithm according to claim 7, characterized in that: The LightGBM algorithm regression prediction model uses the mean absolute error MAE scoring standard. Among them, F is the predicted value, O is the observed value, i is a certain prediction time, and n is the sum of all prediction times.

Images