WO2021077729A1

WO2021077729A1 - Lightning prediction method

Info

Publication number: WO2021077729A1
Application number: PCT/CN2020/090434
Authority: WO
Inventors: 方玉河; 李健; 王钊; 陈玥; 吴大伟; 陶汉涛; 许远根; 陈扬; 张磊; 林卿; 姜志博; 高攀; 李旺
Original assignee: 国网电力科学研究院武汉南瑞有限责任公司
Priority date: 2019-10-23
Filing date: 2020-05-15
Publication date: 2021-04-29
Also published as: CN110796299A; AU2020372283A1

Abstract

A lightning prediction method, comprising the following steps: obtaining basic meteorological parameters of an area to be predicted; calculating high-order meteorological parameters related to lightning based on the high-order meteorological parameters of said area; obtaining lightning positioning observation data of said area, and performing grid processing on the lightning positioning observation data; based on a random forest algorithm, calculating the degree of correlation between each high-order meteorological parameter and lightning, and selecting the high-order meteorological parameter with the highest degree of correlation to lightning; establishing a forecasting model by utilizing an XGBoost algorithm based on the forecasting timeliness and the forecasting time; and based on the high-order meteorological parameters of said area, predicting the spatial distribution and occurrence probability of lightning by utilizing the prediction model.

Description

A lightning prediction method

Technical field

The invention relates to the technical field of disaster prevention and reduction, in particular to a lightning prediction method.

Background technique

Thunder and lightning are often accompanied by lightning and thunder. It is also called lightning. It is a very spectacular and extremely destructive natural phenomenon. The location of thunder and lightning is mostly in the cumulonimbus with intense convection process, or between the electrified thundercloud and the ground protrusion. The occurrence and development of the lightning process is the result of the combined effect of many natural and physical conditions such as atmospheric motion and the earth's magnetic field. As a strong discharge phenomenon, the current value during the occurrence of lightning can reach tens of thousands of amperes. Moreover, the instantaneous voltage of lightning is also very high, reaching several million volts. Therefore, the power of a mid-to-low intensity thunderstorm can reach about 10 million watts, which is equivalent to the output power of a small nuclear power plant. Therefore, the energy released by lightning is huge and its instantaneous destructiveness is extremely strong, which has also attracted widespread attention. It was listed as one of the ten most serious natural disasters in the "United Nations International Decade for Disaster Reduction". Therefore, in order to effectively reduce the impact of lightning disasters on economic and social development and avoid the occurrence of heavy casualties and economic loss accidents, it is very important to carry out lightning early warning.

Lightning warning is an indispensable part of the country's disastrous weather forecast, improving its accuracy and forecasting service level, which is closely related to the development of the whole society and the safety of various industries and people's lives. Common lightning forecasting and early warning methods are mainly radar data extrapolation, direct forecasting with numerical models, empirical forecasts based on meteorological elements, and short-term forecasts based on atmospheric electric field instruments. Among them: direct forecasting with numerical models has high accuracy, but requires The computing power is very large and the cost is very high; the calculation amount required for the extrapolation of radar data and the empirical forecast method based on meteorological elements is far smaller than the numerical model, but the accuracy rate is low; methods such as short-term forecast based on atmospheric electric field instrument The forecast result is more accurate, but the forecast time effect is very small.

The existing lightning early warning methods have disadvantages such as low accuracy, too large required computing resources, and too small forecasting timeliness. How to reduce the trial use of computing power, save costs, improve forecast timeliness, and achieve better accuracy are currently problems that need to be resolved.

Summary of the invention

The purpose of the present invention is to provide a lightning prediction method, which has a small amount of calculation, low cost and high forecast accuracy.

In order to achieve the purpose of the present invention, the technical solutions adopted by the present invention are specifically as follows:

A lightning prediction method includes the following steps:

S1: Obtain the basic meteorological parameters of the area to be predicted;

S2: Calculate the high-order meteorological parameters related to lightning based on the high-order meteorological parameters of the area to be predicted;

S3: Obtain the lightning positioning observation data of the area to be predicted, and grid the lightning positioning observation data;

S4: Based on the random forest algorithm, calculate the correlation degree of each high-order meteorological parameter with thunder and lightning, and select the high-order meteorological parameter with a high degree of correlation with thunder and lightning;

S5: Use XGBoost algorithm to establish a forecast model based on forecast timeliness, forecast times, and high-order meteorological parameters that are highly correlated with lightning;

S6: Based on the high-order meteorological parameters of the area to be predicted, use the forecast model to predict the spatial distribution and occurrence probability of lightning.

As a preference of the above solution, the basic meteorological parameters include temperature, humidity, dew point, vorticity, air pressure, convective precipitation, non-convective precipitation, convective effective potential energy, and radar reflectivity at different altitudes in the area to be predicted.

As a preference of the above solution, the high-level meteorological parameters include A index, K index, Sabouraud index, and strong weather threat index.

As a preference of the above solution, in step S3, gridding the lightning positioning observation data refers to using a grid method to convert the lightning positioning observation data into a grid with the same longitude, latitude, and resolution as the basic meteorological parameters. Grid data.

As a preference of the above solution, in step S4, based on the random forest algorithm, the specific method for calculating the correlation degree of each high-order meteorological parameter with lightning is: taking each high-order meteorological parameter as the feature vector and using the gridded lightning The positioning observation data is used as the target vector to establish a random forest model, and then the outer bag function is used as an evaluation index to calculate the importance of each feature vector, and determine the degree of correlation between high-order meteorological parameters and lightning according to the importance of each feature vector.

As a preference of the above scheme, in step S5, based on the forecast timeliness, forecast times, and high-order meteorological parameters that are highly correlated with lightning, the XGBoost algorithm is used to establish a forecast model as follows: For each high-order meteorological parameter, use high-order meteorological parameters The historical data of the parameters is the feature vector, the grid-processed lightning location historical observation data is the target vector, the linear regression function is the objective parameter, and the hyperopt algorithm is used to perform Bayesian adjustment of the hyperparameters in the XGBoost algorithm to construct The forecast model of high-order meteorological parameters and lightning data at different forecast times is a multi-time forecast model.

As a preference of the above solution, based on the high-order meteorological parameters of the area to be predicted, using a forecast model to predict the spatial distribution and occurrence probability of lightning includes:

(1) Input the high-order meteorological parameters of each forecast time into the multi-time forecast model to obtain the lightning forecast data of each forecast time;

(2) The lightning forecast data sequence of the same forecast time is recombined to generate gridded lightning forecast data based on the high-order meteorological parameters of the area to be predicted.

Compared with the prior art, the present invention has the following beneficial effects:

Compared with numerical forecasting, the lightning forecasting method disclosed by the present invention significantly reduces the calculation amount and greatly reduces the calculation cost; moreover, it uses random forest algorithm and XGBoost algorithm to establish a forecast model. Compared with the numerical forecasting model based on complex fluid mechanics equations, which is easily solved by several Tflop/s, this method has the advantages of small calculation amount and low cost; moreover, compared with the traditional linear model-based meteorological statistical model, it introduces It has more nonlinearity and higher complexity, so the accuracy is higher and its forecast time is equal to the input global model forecast time, up to more than ten days.

The above description is only an overview of the technical solution of the present invention. In order to understand the technical means of the present invention more clearly, it can be implemented in accordance with the content of the specification, and in order to make the above and other objectives, features and advantages of the present invention more obvious and understandable. In the following, preferred embodiments are specifically cited, which are described in detail below.

Detailed ways

In order to further explain the technical means and effects of the present invention to achieve the intended purpose of the invention, the specific implementation, structure, features and effects of the present invention will be described in detail below in conjunction with preferred embodiments:

The invention discloses a lightning prediction method, which includes the following steps:

S1: Obtain the basic meteorological parameters of the area to be predicted.

S4: Based on the random forest algorithm, calculate the degree of correlation between high-order meteorological parameters and lightning, and select high-order meteorological parameters with high degree of correlation with lightning. This is because when the random forest algorithm is used to judge the importance of high-order meteorological parameters, There is no need to consider whether the high-level meteorological parameters are linearly separable, and there is no need to normalize or standardize features;

S5: Use the XGBoost algorithm to establish a forecast model based on the forecast timeliness, forecast times, and high-order meteorological parameters that are highly correlated with lightning.

The XGBoost algorithm is one of the boosting algorithms, and the idea of the Boosting algorithm is to integrate many weak classifiers to form a strong classifier. Moreover, since XGBoost is a boosted tree model, it integrates many tree models in At the same time, a strong classifier is formed. In the present invention, the objective function of lightning forecast is a linear regression function. For each forecast time and each time period, Bayesian optimization method is used to determine the maximum depth, tree Optimize the coefficients such as the number, learning rate, sampling number, and the minimum sample proportion of the end node, and then every period of time, the new observation data obtained is put into the training sample, and the training is retrained to obtain a new forecast model. Therefore, in the present invention, the forecasting effect of the forecasting model can be continuously improved.

As a further preferred solution, the basic meteorological parameters include temperature, humidity, dew point, vorticity, air pressure, convective precipitation, non-convective precipitation, convective effective potential energy, and radar reflectivity at different altitudes in the area to be predicted, specifically , Obtain the 72-hour, 3-hour-by-three-hour forecast of the temperature, humidity, dew point, vorticity and other variables of each pressure layer from the EC global forecast model, and obtain the ground convective precipitation, non-convective precipitation, convective effective potential energy and other variables; Obtain radar reflectivity and so on in the forecast mode.

The high-level meteorological parameters include A index, K index, Sabouraud index and strong weather threat index, among which:

(1) The calculation formula of A index is:

A=T850-T500-(T850-Td850)-(T700-Td700)-(T500-Td500);

(2) The formula for calculating K index is:

The K index is defined as: K=T850-T500+Td850-(T700-Td700);

(3) The Sabouraud index is defined as:

SI=T500-T', where: T'is the temperature of the air block when the moist air mass on the isobar of 850hPa rises along the dry insulation line, and after reaching the condensation height, it rises to 500hPa along the moist insulation line.

(4) The strong weather threat index is defined as:

SWEA=12*Td850+20*(TT-49)+4*WF850+2*WF500+125*(sin(WD500-WD850)+0.2), where: TT is the total index value, if the sub-item of the formula is less than 0, does not count this sub-item, that is, the value is 0, WF is in "m/s" as the unit, the rightmost sub-item must satisfy WD850 at 130°～250°, WD500 at 210°～310°, WD500 is greater than WD850, Calculate when both WF850 and WF500 are greater than 7.5m/s, otherwise it is 0.

It should be noted that in the above definition, T stands for temperature, Td stands for potential temperature, WF stands for wind speed, WD stands for wind direction, and the value of the suffix stands for the pressure layer where the variable is located.

In step S3, gridding the lightning location observation data refers to using the grid method to convert the lightning location observation data into grid data with the same longitude, latitude and resolution as the basic meteorological parameters. This is Because the lightning positioning observation data is station data, the gridding method can be used to convert the lightning positioning observation data into grid data with the same longitude, the same latitude, and the same resolution as the basic meteorological parameters.

In order to better illustrate the method of gridding, take a grid point in the area to be predicted as an example, taking the center of the grid point and R as the radius, the number of lightning strikes is N, when R<20km, And N/R2≥1/(5*5), the grid point value is considered to be 1, otherwise it is 0, and finally a two-dimensional matrix is obtained, which is the gridded lightning data.

In addition, to select high-level meteorological parameters that are highly correlated with lightning is to select one or more parameters from the A index, K index, Sabouraud index, and strong weather threat index to use them as an important reference variable for lightning forecasting.

As a preference of the above scheme, in step S5, based on the forecast timeliness, forecast times, and high-order meteorological parameters that are highly correlated with lightning, the XGBoost algorithm is used to establish a forecast model as follows: For each high-order meteorological parameter, use high-order meteorological parameters The historical data of the parameters is the feature vector, the grid-processed lightning location historical observation data is the target vector, the linear regression function is the objective parameter, and the hyperopt algorithm is used to perform Bayesian adjustment of the hyperparameters in the XGBoost algorithm to construct The forecast model of high-order meteorological parameters and lightning data at different forecast times is the multi-time forecast model, which is specifically: (1) For each high-order meteorological parameter, the grid-processed lightning positioning observation data Historical data is the target vector, linear regression function is the objective parameter, and the hyperopt algorithm is used to perform Bayesian adjustment of the hyperparameters in the XGBoost algorithm such as the number of iterations, the number of trees, and the depth of the tree; (2) In each forecast At the time, the forecast model of high-order meteorological parameters and lightning data can be established, and then a multi-time forecast model can be obtained.

It should be noted that in the present invention, the XGBoost algorithm is used to establish the forecast model because the XGBoost algorithm has the following advantages: (1) The XGBoost algorithm supports linear classifiers, which is equivalent to the introduction of L1 and L2 regularization terms in logistic regression (classification problem) And linear regression (regression problem); (2) The XGBoost algorithm does a second-order Taylor expansion of the cost function, and introduces the first-order derivative and the second-order derivative, so that we can clearly understand what the whole goal is, and step by step Deduced how to learn the tree; (3) When the sample has missing values, XGBoost can automatically learn the splitting direction; (4) XG Boost draws on the approach of RF and supports column sampling, which can not only prevent overfitting, but also Reduce the amount of calculation; (5) The cost function of the XGBoost algorithm introduces a regularization term to control the complexity of the model. The regularization term includes the number of all leaf nodes, and the square sum of the L2 modulus of the score output by each leaf node. From the perspective of Bayesian variance, the regular term reduces the variance of the model and prevents the model from overfitting; (6) XGBoost allocates the learning rate to the leaf nodes after each iteration, reduces the weight of each tree, and reduces each tree. The influence of the tree provides a better learning space for the following; (7) XGBoost tool supports parallelism, but it is not the granularity of the tree, but the granularity of the feature. The most time-consuming step of the decision tree is to sort the value of the feature. XGBoost is Before iteration, pre-sort and save it as a block structure. In each iteration, the structure is reused, which reduces the calculation of the model; the block structure also provides the possibility of parallelism for the model. When splitting the nodes, calculate each feature Select the feature with the largest gain to split in the next step, then the gain of each feature can be performed in multiple threads; (8) Parallel approximate histogram algorithm, when the tree node is split, the gain of each node needs to be calculated If the amount of data is large, sort the features of all nodes to obtain the optimal segmentation point. This greedy method is extremely time-consuming. At this time, the approximate histogram algorithm is introduced to generate efficient segmentation points, that is, split A certain value after subtracting a certain value before splitting to obtain a gain. In order to limit the growth of the tree, a threshold is introduced. When the gain is greater than the threshold, the split is performed. Generally speaking, XGBoost is the most commonly used and one of the most effective models for machine learning modeling of structured data.

As a further solution, based on the basic meteorological parameters of the area to be predicted, based on the high-order meteorological parameters of the area to be predicted, using a forecast model to predict the spatial distribution and occurrence probability of lightning includes:

(2) Recombine the lightning forecast data sequence of the same forecast time to generate gridded lightning forecast data.

The foregoing embodiments are only preferred embodiments of the present invention, and cannot be used to limit the scope of protection of the present invention. Any insubstantial changes and substitutions made by those skilled in the art on the basis of the present invention belong to the present invention. The scope of protection required.

Claims

A lightning prediction method is characterized in that it comprises the following steps:

S1: Obtain the basic meteorological parameters of the area to be predicted;

S2: Calculate the high-order meteorological parameters related to lightning based on the high-order meteorological parameters of the area to be predicted;

S3: Obtain the lightning positioning observation data of the area to be predicted, and grid the lightning positioning observation data;

S4: Based on the random forest algorithm, calculate the correlation degree of each high-order meteorological parameter with thunder and lightning, and select the high-order meteorological parameter with a high degree of correlation with thunder and lightning;

S5: Use XGBoost algorithm to establish a forecast model based on forecast timeliness, forecast times, and high-order meteorological parameters that are highly correlated with lightning;

S6: Based on the high-order meteorological parameters of the area to be predicted, use the forecast model to predict the spatial distribution and occurrence probability of lightning.
The lightning prediction method according to claim 1, wherein the basic meteorological parameters include temperature, humidity, dew point, vorticity, air pressure, convective precipitation, non-convective precipitation, and convective effectiveness at different altitudes in the area to be predicted. Potential energy and radar reflectivity.
The lightning prediction method according to claim 1, wherein the high-level meteorological parameters include A index, K index, Sabouraud index, and strong weather threat index.
The lightning prediction method according to claim 1, characterized in that, in step S3, gridding the lightning positioning observation data refers to using a gridding method to convert the lightning positioning observation data to the basic meteorological parameters. Grid data with the same longitude, latitude, and resolution.
The lightning prediction method according to claim 1, characterized in that, in step S4, based on the random forest algorithm, the specific method for calculating the degree of correlation between each high-order meteorological parameter and lightning is: taking each high-order meteorological parameter as a feature vector , Establish a random forest model with the grid-processed lightning location observation data as the target vector, and then use the outer bag function as the evaluation index to calculate the importance of each feature vector, and determine the height of each feature vector according to the importance of each feature vector The degree of correlation between first-order meteorological parameters and lightning.
The lightning prediction method according to claim 1, characterized in that, in step S5, based on the forecast timeliness, forecast time, and high-order meteorological parameters that are highly correlated with lightning, the XGBoost algorithm is used to establish a forecast model as follows: High-order meteorological parameters, using the historical data of high-order meteorological parameters as the feature vector, the grid-processed lightning location historical observation data as the target vector, and the linear regression function as the objective parameter. The hyperopt algorithm is used to compare the super in the XGBoost algorithm. The parameters are adjusted by Bayesian parameters, and the forecast model of high-order meteorological parameters and lightning data at different forecast times is constructed, that is, a multi-temporal forecast model is obtained.
The lightning prediction method according to claim 6, characterized in that, based on the high-order meteorological parameters of the area to be predicted, using a forecast model to predict the spatial distribution and occurrence probability of lightning comprises:

(1) Input the high-order meteorological parameters of each forecast time into the multi-time forecast model to obtain the lightning forecast data of each forecast time;

(2) Recombine the lightning forecast data sequence of the same forecast time to generate gridded lightning forecast data.