CN111832828B - Intelligent precipitation prediction method based on Fengyun-4 meteorological satellite - Google Patents

Abstract

According to the intelligent precipitation prediction method based on the weather satellite No. four of the cloud, time and space continuity integration is carried out on observed result data of the weather satellite No. four of the cloud, a GBDT mathematical model for precipitation prediction is designed, an initial value of parameters of the GBDT algorithm model is optimized by utilizing a genetic algorithm before the integrated satellite observation data trains the GBDT algorithm model, an optimal parameter initial value of the GBDT algorithm model is obtained, the integrated data trains a precipitation prediction model, and precipitation prediction is completed by utilizing the trained prediction model. According to the invention, through integrating training data, the data of the cloud satellite No. four is used to the maximum extent, the sufficiency and the reality of sample data are ensured, the initial value of the parameters of the GBDT algorithm model is optimized by utilizing the genetic algorithm, the data processing capacity of the prediction model is improved, and the accuracy of the precipitation prediction result and the stability of prediction calculation are improved.

Description

Intelligent precipitation prediction method based on Fengyun-4 meteorological satellite

Technical Field

The present invention belongs to the technical field of meteorological forecasting, and in particular relates to an intelligent precipitation forecasting method based on the Fengyun-4 meteorological satellite.

Background Art

When the FY-4 satellite imager observes, it must take into account cloud guide wind production, flywheel unloading and maximizing the scanning area frequency. Therefore, the observation time range is 5 minutes and 15 minutes, and the observation space range is the entire disk and the Chinese region. It changes many times every day, and there are also vacancies due to flywheel unloading. Therefore, in order to maximize the use of FY-4 satellite data, the observation results of the FY-4 satellite must be continuously integrated in time and space.

However, the mathematical model of the existing precipitation prediction method based on the Fengyun-4 meteorological satellite cannot integrate the observation result data of the Fengyun-4 satellite in time and space, resulting in poor robustness of the mathematical model trained with the observation result data of the Fengyun-4 satellite, poor accuracy of the prediction results, and low credibility. In addition, the existing mathematical model for precipitation prediction usually adopts a single algorithm and cannot integrate and optimize multiple effective algorithms, resulting in insufficient data processing ability of the mathematical model itself, and poor adaptability, inaccurate prediction results, and insufficient stability for specific precipitation predictions.

Summary of the invention

The purpose of the present invention is to solve the above-mentioned technical problems existing in the prior art and to provide an intelligent precipitation prediction method based on the Fengyun-4 meteorological satellite, which uses the observation data of the Fengyun-4 satellite to complete the prediction of precipitation in the Chinese region.

The specific technical solutions of the present invention are as follows:

Step 1: Time integration of satellite observation data:

The average values of the satellite's conventional observation data for China in each channel in three consecutive 5-minute time periods are integrated into a 15-minute data, and 15 minutes is set as a time period. That is, the values of each channel of the newly generated 15-minute observation data for China are:

Where i is the channel number of FY-4, ranging from 1 to 14, j is the sequence number of the 5-minute data, and n is 3, which is the number of 5-minute data to be averaged.

For the missing data at 17:15 due to the stop of observation for flywheel unloading at midnight, the data at the satellite descent point is replaced by the average of the data at 17:00 and 17:30. Since the data at 17:00 is the full disk map observation and the data at 17:30 is the China regional observation, the ranges are different and need to be integrated into the China regional observation size before the average calculation;

Step 2: Spatial integration of satellite scanning data:

The full-disk conventional observation data after time integration is cut to obtain the same area as the conventional observation in the Chinese region, so that all data have the same range and time frequency. The time and space integration of satellite data is completed. The integrated satellite data is 4 times an hour, and the range is the conventional observation range of the Chinese region;

Step 3: Fengyun-4 satellite precipitation prediction channel selection

Fengyun-4 satellite has a total of 14 channels. When satellites predict precipitation, they will select an infrared channel and a water vapor channel for calculation according to the physical mechanism. When more channels are selected, it is impossible to analyze the relationship between the channels and their role in precipitation prediction. However, each channel contains information that contributes to precipitation. If fewer channels are selected, information will be lost, and thus the accuracy of precipitation cannot be improved. In order to fully mine the information in the channels, when using Fengyun-4 satellite data, 8 satellite channels from 7 to 14 in the list are used;

Step 4: Genetic algorithm settings and parameter configuration

The mathematical model combining genetic algorithm and GBDT (Gradient Boosting Decision Tree) algorithm is used to predict precipitation. The initial values of the parameters of the GBDT algorithm model are optimized using the genetic algorithm to obtain the optimal initial values of the GBDT algorithm model.

Among them, the specific design of the genetic algorithm is:

1. Population initialization and chromosome encoding

The individual encoding method is real number encoding. Each individual is a real number string. The individual contains all the parameters of the GBDT algorithm model. The number of initial populations is: 20-100;

2. Determine the objective function and fitness function

According to all the parameters of the GBDT algorithm model obtained by the individual, the system output is predicted after the GBDT algorithm model is trained with the training data. The absolute value of the error between the expected output and the actual output is taken as the individual fitness value F, and its calculation formula is:

Where n is the number of network output nodes; _yi is the expected output of the leaf node of the i-th regression tree of the GBDT algorithm model; _oi is the actual output of the leaf node of the i-th regression tree; k is the coefficient;

3. Select an operation

The selection operation is set as the proportional selection method, that is, the selection strategy based on the fitness ratio, and the selection probability pi of each individual i is:

In the formula, Fi is the fitness value of individual i; N is the number of individuals in the population;

4. Crossover Operation

Since individuals are coded with real numbers, the crossover operation method uses the real number crossover method. The crossover operation method of the kth chromosome _ak and the lth chromosome a _l in the jth gene is as follows:

a _kj = a _kj (1-b) + a _lj b

a _lj = a _lj (1-b) + a _kj b

Where, coefficient b is a random number between [0,1]; the crossover probability is set to 0.4-0.99;

5. Mutation Operation

Select the jth gene _aij of the i-th chromosome individual for mutation operation, the method is as follows:

In the formula, a _max is the upper bound of gene a _ij ; a _min is the lower bound of gene a _ij ; g is the current iteration number, the termination evolution number of the genetic algorithm is 100-500; G _max is the maximum evolution number; r is a random number between [0,1]; the mutation probability is 0.0001-0.1;

Step 5: Setting and calculation process of GBDT algorithm

In step 4, the initial values of the parameters of the GBDT algorithm model are optimized by using a genetic algorithm to obtain the optimal initial values of the parameters of the GBDT algorithm model; in step 5, the historical data scanned by the integrated satellite observation point, which may be the latitude and longitude information, the annual serial number of the day and the altitude of the point, are substituted into the GBDT algorithm model to obtain the predicted precipitation corresponding to the above-mentioned satellite observation point; the predicted precipitation is compared with the actual precipitation in the historical data scanned by the satellite observation point to obtain the residual, and the GBDT algorithm model is further trained; the optimal parameter selection of the GBDT algorithm model is obtained through training, and then the real-time data of the observation point scanned by the satellite in real time, such as the longitude and latitude information, the annual serial number of the day and the altitude of the point, are substituted into the trained GBDT algorithm model to obtain the corresponding precipitation prediction value of the satellite observation point;

(1) The GBDT algorithm is set as:

① Choice of loss function

For the GBDT algorithm to predict precipitation, it is necessary to find the appropriate parameters of the GBDT algorithm and measure whether they are verified parameters of the GBDT algorithm. The loss function is used to determine the parameters. The loss function is:

L(y,F)＝|yF|， Where y is the actual value and F(x) is the predicted value. Huber's loss Obtaining the most appropriate parameters for each regression tree is equivalent to minimizing the loss function L;

②Objective function

The objective function is to obtain the most appropriate parameters for each regression tree, that is, the allocation ratio _ρt of the tth regression tree and the parameters of the tth regression tree: Where ρ is the allocation ratio of each regression tree, that is, the learning rate or step size of the tree, and θ is the parameter of the regression tree;

The parameter iteration

The step size is selected

Where Ex _,y is the error rate of each regression tree iteration, x represents the input data, f( _xi ) represents the predicted value of the i-th iteration, and h represents the second-order derivative. h(x,θ) represents the regression tree, and h( _xi , _θt ) represents the i-th iteration of the t-th regression tree.

③ Calculation of residuals:

When the GBDT algorithm trains the regression tree, it needs to use the residual as the target value to train the next regression tree until the number of established regression trees reaches the requirement and the residual _rit reaches the expected range. The training is stopped. _rit also represents the direction of the weak learner established in the i-th iteration.

(2) The iterative process of the GBDT algorithm is:

Each decision tree of GBDT is trained based on the errors in the classification results of the previous decision tree. Each calculation is to reduce the residual of the previous one. In order to eliminate the residual, GBDT performs multiple iterations. Each iteration builds a new model in the gradient direction of residual reduction. The specific iteration process is as follows:

① Initialize the first equation And the initial value θ ₀ of the parameter of the GBDT algorithm model is optimized by using a genetic algorithm to obtain the optimal initial value of the parameter of the GBDT algorithm model;

②The residual is calculated at each iteration R _it also represents the direction of the weak learner established in the i-th iteration to train the next regression tree;

③ Construct a regression equation {( _xi , _rit )} _i=1,...,n with the residual as the target, _wherexi is the imported data, which can be the latitude and longitude information of the satellite observation point, the year serial number of the day, and the altitude of the point;

④Find the most suitable parameter selection at present:

⑤After M iterations, the final prediction result is obtained: _yi is the predicted precipitation corresponding to the above satellite observation point.

Training sample settings: The traditional precipitation model uses one template for the whole country, which can neither handle the precipitation differences caused by different altitudes in different regions, nor the precipitation differences caused by different longitudes and latitudes, nor the precipitation differences caused by different seasons. The present invention matches the actual precipitation product and the satellite resolution to after 5 kilometers. In order to make the prediction more stable, the data of 9 points of 3×3 around the observation point is innovatively used, and the longitude and latitude information, the year serial number of the day and the altitude of the point are added as the input of a sample. The input of this feature innovatively takes into account time, location and altitude at the same time, which is a great improvement over the previous unified model for precipitation across the country. For an input training sample, a total of 292 features are included,

Training process: In order to speed up the training time and reduce the prediction time, multiple models are trained by region. The China regional observation range is 3°-55° north latitude and 60°-137° east longitude. A model is trained for each data within the range of 4°×4°, so that parallel calculations can be performed during training and prediction, which improves efficiency and greatly shortens the prediction time. The span in the latitude direction is 55°-3°=52°, and a total of 13 regional models are required. The span in the longitude direction is 137°-60°=77°, and a total of 20 models are required. The 20th model only requires 1° longitude × 4° latitude data. In other words, a total of 20×13=260 regional models need to be trained. For each model, input the data from 2019 for training. The satellite resolution is 0.04°, so there are 100×100 points in each standard area, multiplied by 365 days, for a total of 3.65 million training data for training.

Forecast settings: For satellite data outside the training set, when the satellite obtains 4-hour data within an hour, it is split into 4°×4°. Each area is assigned to the corresponding model for prediction, and finally the prediction results are spliced together to form the precipitation distribution in China.

The precipitation prediction method based on the Fengyun-4 satellite of the present invention has achieved the following technical effects:

1. The observation result data of the Fengyun-4 satellite are continuously integrated in time and space, and the data of the Fengyun-4 satellite are used to the maximum extent to ensure the sufficiency and authenticity of the sample data. The integrated data are used to train the precipitation prediction model to improve the accuracy and credibility of the results of the precipitation prediction method.

In view of the data characteristics of the observation results of the Fengyun-4 satellite, a mathematical model for precipitation prediction is involved. The genetic algorithm and the GBDT algorithm model are deeply integrated and optimized. Before the integrated satellite observation data are used to train the GBDT algorithm model, the initial values of the parameters of the GBDT algorithm model are optimized using the genetic algorithm to obtain the optimal initial values of the parameters of the GBDT algorithm model. Based on the initial values of the parameters, the iterative calculation of the GBDT algorithm model is further completed. Compared with the random generation of the initial values of the GBDT algorithm model parameters in the prior art, the data processing capability of the mathematical model is improved, and the accuracy of the precipitation prediction results and the stability of the prediction calculation are improved.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1: GBDT algorithm flow chart.

DETAILED DESCRIPTION

In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are 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.

When the FY-4 satellite imager is observing, it must take into account cloud guide wind production, flywheel unloading and maximizing the scanning area frequency. Therefore, the observation time is 5 minutes and 15 minutes, and the spatial range of observation is the entire disk and the Chinese area. It changes many times every day. In addition, due to the vacancies caused by flywheel unloading, in order to maximize the use of FY-4 satellite data, the observation results of the FY-4 satellite must be integrated continuously in time and space.

1. Satellite scanning data time integration

The observation mode of the FY-4 satellite imager is as follows: the imager obtains 40 full disk cloud images and 165 China regional cloud images every day. The China regional observation range is 3°-55°N and 60°-137°E. Full disk observation is performed once every hour, from the hour to 15 minutes after the hour.

Three consecutive full disk observations are conducted every three hours, with the full disk cloud images at 00:00 (UTC, the same below)/03:00/06:00/09:00/12:00/15:00/18:00/21:00 and one full disk cloud image before and after the time. For example, the three consecutive full disk observations at 00:00 are at 23:45-23:59:59, 00:00-00:14:59, and 00:15-00:29:59.

When there is no full disk observation, the observation of China area will be carried out for 5 minutes. Positioning and calibration observation will be carried out in every 15-minute observation interval. The flywheel will be unloaded and observation will be stopped at midnight at the subsatellite point. The time is 17:15-17:30.

Since the observation time of satellite data is irregular, it is not conducive to the input of the later algorithm. Therefore, the satellite data is integrated, and the average of all channels of three consecutive 5-minute routine observations in the Chinese region are integrated into a 15-minute data. That is, the channel values of the newly generated 15-minute observation data in the Chinese region are:

Where i is the channel number of FY-4, ranging from 1 to 14, j is the sequence number of the 5-minute data, and n is 3, which is the number of 5-minute data to be averaged.

For the missing data at 17:15, which was stopped for flywheel unloading at midnight, the average of the two hours before and after was used to replace it. Since the data at 17:00 was for full disk observation and the data at 17:30 was for China observation, the ranges were different and needed to be integrated into the China observation size before the average calculation.

2. Spatial integration of satellite scanning data

The full-disk routine observation data after time integration is cut to obtain the same area as the routine observation in the Chinese region, so that all data have the same range and time frequency. The time and space integration of satellite data is completed. The integrated satellite data is collected 4 times an hour, and the range is the routine observation range of the Chinese region.

3. Fengyun-4 satellite precipitation prediction channel selection

Fengyun-4 satellite has a total of 14 channels, as shown in Table 1. Generally, when satellites predict precipitation, an infrared channel plus a water vapor channel will be selected for calculation according to the physical mechanism, because when more channels are selected, the relationship between the channels and the effect on precipitation prediction cannot be analyzed. However, each channel contains information that contributes to precipitation. If fewer channels are selected, information loss will occur, thereby failing to improve the accuracy of precipitation. The present invention uses artificial intelligence algorithms to predict precipitation, which can better mine the information in the channels. Therefore, when using Fengyun-4 satellite data, 8 satellite channels 7-14 in the list are used.

Table 1 List of FY-4 satellite channels

Channel Number Center wave number (μm) Channel Type 1 0.47 Visible light 2 0.65 Visible light 3 0.83 Near Infrared 4 1.37 Shortwave infrared 5 1.61 Shortwave infrared 6 2.22 Shortwave infrared 7 3.72H Mid-wave infrared 8 3.72L Mid-wave infrared 9 6.25 Water vapor 10 7.1 Water vapor 11 8.5 Long wave infrared 12 10.8 Long wave infrared 13 12 Long wave infrared 14 13.5 Long wave infrared

4. Algorithm selection and parameter configuration

GBDT gradient boosting decision tree is an iterative decision tree algorithm, which is based on the forward distribution algorithm and the additive model, and is composed of multiple decision trees. It is an integrated algorithm that realizes classification and regression by reducing the residuals generated in the learning process, and finally adds up the conclusions of all trees to make decisions. There are two types of trees, namely classification trees and regression trees. Classification trees are generally used to deal with classification problems, and regression trees are generally used to deal with prediction problems. GBDT is an algorithm with strong generalization ability. Since the core of the GBDT algorithm is to fit the residuals generated by the previous tree through each tree and accumulate the results of all trees through a series of formula calculations as the final prediction output, and the classification tree is not easy to implement the above process, so the present invention uses the tree in the GBDT algorithm as a regression tree. The flow chart is shown in Figure 1.

The genetic algorithm is used to optimize the initial values of the parameters of the GBDT algorithm model to obtain the optimal initial values of the parameters of the GBDT algorithm model; then the historical data scanned by the integrated satellite observation point, such as the longitude and latitude information, the annual serial number of the day and the altitude of the point, are substituted into the GBDT algorithm model to obtain the predicted precipitation corresponding to the above-mentioned satellite observation point; the predicted precipitation is compared with the actual precipitation in the historical data scanned by the satellite observation point to obtain the residual, and the GBDT algorithm model is further trained; the optimal parameter selection of the GBDT algorithm model is obtained through training, and then the real-time data of the observation point scanned by the satellite in real time, such as the longitude and latitude information, the annual serial number of the day and the altitude of the point, are substituted into the trained GBDT algorithm model to obtain the corresponding precipitation prediction value of the satellite observation point.

Among them, the settings and parameter configuration of the genetic algorithm:

The mathematical model combining genetic algorithm and GBDT algorithm is used to predict precipitation; the initial values of the parameters of the GBDT algorithm model are optimized by genetic algorithm to obtain the optimal initial values of the GBDT algorithm model;

Among them, the specific design of the genetic algorithm is:

1. Population initialization and chromosome encoding

The individual encoding method is real number encoding. Each individual is a real number string. The individual contains all the parameters of the GBDT algorithm model. The number of initial populations is: 20-100;

2. Determine the objective function and fitness function

According to all the parameters of the GBDT algorithm model obtained by the individual, the system output is predicted after the GBDT algorithm model is trained with the training data. The absolute value of the error between the expected output and the actual output is taken as the individual fitness value F, and its calculation formula is:

Where n is the number of network output nodes; _yi is the expected output of the leaf node of the i-th regression tree of the GBDT algorithm model; _oi is the actual output of the leaf node of the i-th regression tree; k is the coefficient;

3. Select an operation

The selection operation is set as the proportional selection method, that is, the selection strategy based on the fitness ratio, and the selection probability pi of each individual i is:

In the formula, Fi is the fitness value of individual i; N is the number of individuals in the population;

4. Crossover Operation

Since individuals are coded with real numbers, the crossover operation method uses the real number crossover method. The crossover operation method of the kth chromosome _ak and the lth chromosome a _l in the jth gene is as follows:

a _kj = a _kj (1-b) + a _lj b

a _lj = a _lj (1-b) + a _kj b

Where, coefficient b is a random number between [0,1]; the crossover probability is set to 0.4-0.99;

5. Mutation Operation

Select the jth gene _aij of the i-th chromosome individual for mutation operation, the method is as follows:

In the formula, a _max is the upper bound of gene a _ij ; a _min is the lower bound of gene a _ij ; g is the current iteration number, the termination evolutionary number of the genetic algorithm is 100-500; G _max is the maximum evolutionary number; r is a random number between [0,1]; and the mutation probability is 0.0001-0.1.

Among them, the setting and calculation process of GBDT algorithm:

In step 4, the initial values of the parameters of the GBDT algorithm model are optimized by using a genetic algorithm to obtain the optimal initial values of the parameters of the GBDT algorithm model; in step 5, the historical data scanned by the integrated satellite observation point, which may be the latitude and longitude information, the annual serial number of the day and the altitude of the point, are substituted into the GBDT algorithm model to obtain the predicted precipitation corresponding to the above-mentioned satellite observation point; the predicted precipitation is compared with the actual precipitation in the historical data scanned by the satellite observation point to obtain the residual, and the GBDT algorithm model is further trained; the optimal parameter selection of the GBDT algorithm model is obtained through training, and then the real-time data of the observation point scanned by the satellite in real time, such as the longitude and latitude information, the annual serial number of the day and the altitude of the point, are substituted into the trained GBDT algorithm model to obtain the corresponding precipitation prediction value of the satellite observation point;

(1) The GBDT algorithm is set as:

① Choice of loss function

For the GBDT algorithm to predict precipitation, it is necessary to find the appropriate parameters of the GBDT algorithm and measure whether they are verified parameters of the GBDT algorithm. The loss function is used to determine the parameters. The loss function is:

L(y,F)＝|yF|， Where y is the actual value and F(x) is the predicted value. Huber's loss Obtaining the most appropriate parameters for each regression tree is equivalent to minimizing the loss function L;

②Objective function

The objective function is to obtain the most appropriate parameters for each regression tree, that is, the allocation ratio _ρt of the tth regression tree and the parameters of the tth regression tree: Where ρ is the allocation ratio of each regression tree, that is, the learning rate or step size of the tree, and θ is the parameter of the regression tree;

The parameter iteration

The step size is selected

Where Ex _,y is the error rate of each regression tree iteration, x represents the input data, f( _xi ) represents the predicted value of the i-th iteration, and h represents the second-order derivative. h(x,θ) represents the regression tree, and h( _xi , _θt ) represents the i-th iteration of the t-th regression tree.

③ Calculation of residuals:

When the GBDT algorithm trains the regression tree, it needs to use the residual as the target value to train the next regression tree until the number of established regression trees reaches the requirement and the residual _rit reaches the expected range. The training is stopped. _rit also represents the direction of the weak learner established in the i-th iteration.

(2) The iterative process of the GBDT algorithm is:

Each decision tree of GBDT is trained based on the errors in the classification results of the previous decision tree. Each calculation is to reduce the residual of the previous one. In order to eliminate the residual, GBDT performs multiple iterations. Each iteration builds a new model in the gradient direction of residual reduction. The specific iteration process is as follows:

① Initialize the first equation And the initial value θ ₀ of the parameter of the GBDT algorithm model is optimized by using a genetic algorithm to obtain the optimal initial value of the parameter of the GBDT algorithm model;

②The residual is calculated at each iteration R _it also represents the direction of the weak learner established in the i-th iteration to train the next regression tree;

③ Construct a regression equation {( _xi , _rit )} _i=1,...,n with the residual as the target, _wherexi is the imported data, which can be the latitude and longitude information of the satellite observation point, the year serial number of the day, and the altitude of the point;

④Find the most suitable parameter selection at present:

⑤After M iterations, the final prediction result is obtained: _yi is the predicted precipitation corresponding to the above satellite observation point.

After training and learning of the training set and 5 cross-validation of the test set, the optimal parameter settings of the GBDT algorithm model are finally obtained as shown in Table 2. Under this parameter, the average precipitation error in 2019 is 0.0476 mm, which is much better than the default parameter error of 0.327 mm.

Table 2 Parameters

5. Training sample settings

The traditional precipitation model uses one template for the whole country, which can neither handle the precipitation differences caused by different altitudes in different regions, nor the precipitation differences caused by different longitudes and latitudes, nor the precipitation differences caused by different seasons. The present invention matches the real-time precipitation product and the satellite resolution to 5 kilometers. In order to make the prediction more stable, the data of 9 points around the observation point in a 3×3 shape is innovatively used, with the longitude and latitude information, the year serial number of the day, and the altitude of the point as the input of a sample. The input of this feature innovatively takes into account time, location and altitude at the same time, which is a great improvement over the previous unified model for precipitation across the country. For an input training sample, a total of 292 features are included, as shown in the following table:

Table 3 Input sample content list

The corresponding label is the precipitation at this point in the hour.

6. Parallel computing settings

In order to speed up the training time and reduce the prediction time, multiple models are trained by region. The Chinese regional observation range is 3°-55° north latitude and 60°-137° east longitude. A model is trained for each data within the range of 4°×4°, so that parallel calculations can be performed during training and prediction, which improves efficiency and greatly shortens the prediction time. The span in the latitude direction is 55°-3°=52°, and a total of 13 regional models are required. The span in the longitude direction is 137°-60°=77°, and a total of 20 models are required, of which the 20th model only requires 1° longitude × 4° latitude data. That is, a total of 20×13=260 regional models need to be trained. For each model, the data of 2019 is input for training, and the satellite resolution is 0.04°. Then in each standard area, there are 100×100 points, multiplied by 365 days, a total of 3.65 million training data for training.

7. Forecast Settings

For satellite data outside the training set, when the satellite obtains 4 time-of-day data within an hour, it is split into 4°×4°. Each area is assigned to the corresponding model for prediction, and finally the prediction results are spliced together to form the precipitation distribution in China.

The precipitation prediction method based on the Fengyun-4 satellite has achieved the following technical results:

1. The observation result data of the Fengyun-4 satellite are continuously integrated in time and space, and the data of the Fengyun-4 satellite are used to the maximum extent to ensure the sufficiency and authenticity of the sample data. The integrated data are used to train the precipitation prediction model to improve the accuracy and credibility of the results of the precipitation prediction method.

2. According to the data characteristics of the observation results of the Fengyun-4 satellite, a mathematical model for precipitation prediction was designed, and the genetic algorithm and the GBDT algorithm model were deeply integrated and optimized. Before the integrated satellite observation data were used to train the GBDT algorithm model, the initial values of the parameters of the GBDT algorithm model were optimized using the genetic algorithm to obtain the optimal initial values of the parameters of the GBDT algorithm model. Based on the initial values of the parameters, the iterative calculation of the GBDT algorithm model was further completed. Compared with the random generation of the initial values of the GBDT algorithm model parameters in the prior art, the data processing capability of the mathematical model was improved, and the accuracy of the precipitation prediction results and the stability of the prediction calculation were improved.

The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims (4)

1. A precipitation prediction method based on Fengyun-4 satellite, which uses the observation data of Fengyun-4 satellite to complete the prediction of precipitation in China, and is characterized by comprising the following steps: Step 1: Time integration of satellite observation data: The average values of the satellite's conventional observation data for China in each channel in three consecutive 5-minute time periods are integrated into a 15-minute data, and 15 minutes is set as a time period. That is, the values of each channel of the newly generated 15-minute observation data for China are: Where i is the channel number of FY-4, ranging from 1 to 14, j is the sequence number of the 5-minute data, and n is 3, which is the number of 5-minute data to be averaged; Step 2: Spatial integration of satellite scanning data: The full-disk conventional observation data after time integration is cut to obtain the same area as the conventional observation in the Chinese region, so that all data have the same range and time frequency. The time and space integration of satellite data is completed. The integrated satellite data is 4 times an hour, and the range is the conventional observation range of the Chinese region; Step 3: Fengyun-4 satellite precipitation prediction channel selection: The data of 8 satellite channels of Fengyun-4 satellite were selected for calculation, and the relationship between the channels and their role in precipitation prediction were analyzed; Step 4: Genetic algorithm settings and parameter configuration: The mathematical model combining genetic algorithm and GBDT algorithm is used to predict precipitation; the initial values of the parameters of the GBDT algorithm model are optimized by genetic algorithm to obtain the optimal initial values of the GBDT algorithm model; Among them, the specific design of the genetic algorithm is: (1) Population initialization and chromosome encoding The individual encoding method is real number encoding. Each individual is a real number string. The individual contains all the parameters of the GBDT algorithm model. The number of initial populations is: 20-100; (2) Determine the objective function and fitness function According to all the parameters of the GBDT algorithm model obtained by the individual, the system output is predicted after the GBDT algorithm model is trained with the training data. The absolute value of the error between the expected output and the actual output is taken as the individual fitness value F, and its calculation formula is: Where n is the number of network output nodes; _yi is the expected output of the leaf node of the i-th regression tree of the GBDT algorithm model; _oi is the actual output of the leaf node of the i-th regression tree; k is the coefficient; (3) Select operation The selection operation is set as the proportional selection method, that is, the selection strategy based on the fitness ratio, and the selection probability _pi of each individual i is: In the formula, _Fi is the fitness value of individual i; _Fj is the fitness value of individual j; N is the number of individuals in the population; (4) Crossover operation Since individuals are coded with real numbers, the crossover operation method uses the real number crossover method. The crossover operation method of the kth chromosome _ak and the lth chromosome a _l in the jth gene is as follows: a _kj = a _kj (1-b) + a _lj b a _lj = a _lj (1-b) + a _kj b Where, coefficient b is a random number between [0,1]; the crossover probability is set to 0.4-0.99; (5) Mutation Operation Select the jth gene _aij of the i-th chromosome individual for mutation operation, the method is as follows: In the formula, a _max is the upper bound of gene a _ij ; a _min is the lower bound of gene a _ij ; g is the current iteration number, the termination evolution number of the genetic algorithm is 100-500; G _max is the maximum evolution number; r is a random number between [0,1]; the mutation probability is 0.0001-0.1; Step 5: Setting and calculation process of GBDT algorithm In step 4, the initial values of the parameters of the GBDT algorithm model are optimized by using a genetic algorithm to obtain the optimal initial values of the parameters of the GBDT algorithm model; in step 5, the historical data scanned by the integrated satellite observation point is substituted into the GBDT algorithm model to obtain the predicted precipitation corresponding to the above satellite observation point; the predicted precipitation is compared with the actual precipitation in the historical data scanned by the satellite observation point to obtain the residual, and the GBDT algorithm model is further trained; the optimal parameter selection of the GBDT algorithm model is obtained through training, and the real-time data of the observation point scanned by the satellite in real time is substituted into the trained GBDT algorithm model to obtain the corresponding precipitation prediction value of the satellite observation point; (1) The GBDT algorithm is set as: ① Choice of loss function For the GBDT algorithm to predict precipitation, it is necessary to find the appropriate parameters of the GBDT algorithm and measure whether they are verified parameters of the GBDT algorithm. The loss function is: L(y,F)＝|yF|, Where y is the actual value and F(x) is the predicted value. Huber's loss Obtaining the most appropriate parameters for each regression tree is equivalent to minimizing the loss function L; ②Objective function The objective function is to obtain the most appropriate parameters for each regression tree, that is, the allocation ratio _ρt of the tth regression tree and the parameters of the tth regression tree: Where ρ is the allocation ratio of each regression tree, that is, the learning rate or step size of the tree, and θ is the parameter of the regression tree; The parameter iteration The step size is selected Where Ex _,y is the error rate of each regression tree iteration, x represents the input data, f( _xi ) represents the predicted value of the i-th iteration, and h represents the second-order derivative. h(x,θ) represents the regression tree, and h( _xi , _θt ) represents the i-th iteration of the t-th regression tree. ③ Calculation of residuals: When the GBDT algorithm trains the regression tree, it needs to use the residual as the target value to train the next regression tree until the number of established regression trees reaches the requirement and the residual _rit reaches the expected range. The training is stopped. _rit also represents the direction of the weak learner established in the i-th iteration. (2) The iterative process of the GBDT algorithm is: Each decision tree of GBDT is trained based on the errors in the classification results of the previous decision tree. Each calculation is to reduce the residual of the previous one. In order to eliminate the residual, GBDT performs multiple iterations. Each iteration establishes a new model in the gradient direction of residual reduction. The specific iteration process is as follows: ① Initialize the first equation And the initial value θ ₀ of the parameter of the GBDT algorithm model is optimized by using a genetic algorithm to obtain the optimal initial value of the parameter of the GBDT algorithm model; ②The residual is calculated at each iteration r _it represents the direction of the weak learner established in the i-th iteration to train the next regression tree; ③ Using the residual as the target, construct the regression equation {( _xi , _rit )} _i=1,...,n , _wherexi is the imported data, _xi is the latitude and longitude information of the satellite observation point, the year serial number of the day, and the altitude of the point; ④Find the most suitable parameter selection at present: ⑤After M iterations, the final prediction result is obtained: _yi is the predicted precipitation corresponding to the above satellite observation point. 2. The precipitation prediction method based on the Fengyun-4 satellite as described in claim 1 is characterized in that the training sample is set with the observation point as the center, and the data of 9 points around 3×3 are added with longitude and latitude information, the year serial number of the day and the altitude of the point as the input of a sample; for one input training sample, a total of 292 features are included. 3. The precipitation prediction method based on the Fengyun-4 satellite as described in any one of claims 1-2 is characterized in that, in order to speed up the training time and reduce the prediction time at the same time, multiple models are trained by region. The Chinese regional observation range is 3°-55° north latitude and 60°-137° east longitude; a model is trained for each data within the range of 4°×4°, the latitude span is 55°-3°=52°, and a total of 13 regional models are required. The longitude span is 137°-60°=77°, and a total of 20 models are required, of which the 20th model only requires 1° longitude × 4° latitude data; that is, a total of 20×13=260 regional models need to be trained; for each model, the data of 2019 is input for training, and the satellite resolution is 0.04°, so in each standard area, there are 100×100 points, multiplied by 365 days, a total of 3.65 million training data are trained. 4. The precipitation prediction method based on the Fengyun-4 satellite as described in any one of claims 1-2 is characterized in that, for satellite data outside the training set, after the satellite obtains 4 time data within one hour, 4°×4° splitting is performed; each area is respectively assigned to the corresponding model for prediction, and finally the prediction results are spliced together to form the precipitation distribution in the Chinese region.