CN114626578A

CN114626578A - Method for forecasting freezing rain by using artificial intelligence

Info

Publication number: CN114626578A
Application number: CN202210158808.7A
Authority: CN
Inventors: 温晗秋子; 万鼎煜; 张平文
Original assignee: Peking University
Current assignee: Peking University
Priority date: 2022-02-21
Filing date: 2022-02-21
Publication date: 2022-06-14
Anticipated expiration: 2042-02-21
Also published as: CN114626578B

Abstract

The invention discloses a method for forecasting freezing rain by artificial intelligence, which comprises the steps of preprocessing rainfall type observation data and converting rainfall phase states; dividing the preprocessed precipitation type observation data into a training data set and a verification data set according to the proportion of 90% to 10%; matching geographic positions according to the observation data time, and analyzing feature data of each atmospheric pressure layer of 500hPa-1000hPa and longitude, latitude and height of each forecast station in the data to construct a feature data set; taking observation data as a label, combining with a characteristic data set, and performing freezing rain precipitation phase state forecasting model training through a LightGBM algorithm to obtain a forecasting model and storing the forecasting model; and directly outputting the forecast result of the freezing rain obtained by forecasting by using the characteristic data set of the forecasting model. According to the invention, by constructing the forecasting model, characteristic engineering processing is carried out on the atmospheric pressure layer characteristic data and the longitude, the latitude and the height of each forecasting station, so that the rapid and accurate forecasting of the freezing rain in winter is realized.

Description

Method for forecasting freezing rain by utilizing artificial intelligence

Technical Field

The invention relates to the technical field of meteorological services, in particular to a method for forecasting freezing rain by using artificial intelligence.

Background

Precipitation is one of key links of earth water circulation and energy circulation (Zhang T J, 2005), precipitation phase states comprise a plurality of forms of rain, snow, rain and snow, freezing rain, ice particles, hail and the like, precipitation in different phase states has huge influence on land surface substances and energy circulation (Wu B Y et al, 2009), influence generated by different phase states with the same precipitation amount is remarkably different (Sunyan and the like, 2014; Wangchun and the like, 2005), wherein the influence of freezing rain is the most serious, and the freezing rain is one of the most dangerous disasters in winter and early spring. The freezing rain is a weather phenomenon that when falling on the ground or exposed objects, the freezing rain is rapidly condensed into ice, the water drops are frozen when being impacted, a layer of ice glaze can be formed and covers any object with the temperature lower than zero degrees centigrade (Richard wheels, 2014), when the rainfall is large and the duration is long, disastrous results are possibly brought, further, the infrastructure and vegetation are damaged, the production life and the natural environment (Call, 2010) of human are influenced, and particularly in winter, the precipitation phase state forecast quality is a core influence factor of the precipitation quantity forecast accuracy. Then, the freezing rain, a special type of precipitation in winter, can cause "black ice" on the road surface, called invisible killer on the road, which can also cause serious damage to the power grid, aviation, outdoor sports in winter, agriculture and the like. In recent years, an increasing tendency of the frequency of freezing rain has been observed in many regions of the world, such as the central south canada, the long island of new york, and the southeast of china. The false reporting and the missing reporting of the freezing rain and other disastrous precipitation in winter can seriously affect and reduce the working efficiency of cities, traffic and power supply systems for coping with the disasters, and increase the corresponding loss. Therefore, the improvement of the freezing rain forecasting capability has great significance for traffic management and road maintenance, aviation ground deicing operation, winter power planning and infrastructure maintenance, disaster loss control and the like.

In addition, as is well known, the precipitation phase state forecast is one of the biggest problems faced by forecasters in winter, so that the forecast of the precipitation phase state in winter starts from the seventies of the twentieth century, the main development methods of the method comprise an experience forecast method based on observation data, a numerical mode forecast based on a physical mechanism, a statistical forecast integrating observation and a numerical mode, and the like, and the method can be mainly divided into three categories, wherein the first category is to establish indexes and regression equations based on observation or numerical weather forecast, and is referred to as an index criterion method; the second method is a micro-physical method and an ensemble forecasting method based on a numerical weather forecasting mode; the third category of methods is artificial intelligence forecasting methods based on application decision trees, artificial neural networks, deep learning, etc. of observed data and numerical weather forecasting products.

The thermodynamic and microscopic physical processes of freezing rain formation are complex and difficult to interpret by a unified mechanism. For example, there may be a large difference in vertical thermal structure of its corresponding high-level atmosphere for the occurrence of surface freezing rain under the same terrain conditions. Taking the Beijing area as an example, according to 53 year calendar history observation data and NCEP reanalysis data analysis, 50% of freezing rain occurs when a warm layer exists in the vertical atmosphere, and the other 50% of freezing rain events do not observe an atmosphere warm layer. Another survey on the aca mountains cold air dam in the united states shows that there can be multiple zero degree layers of vertical atmosphere in the event of freezing rain. In addition, freezing rain may also occur in situations where the vertical atmosphere is well below freezing, in which case the temperature of the supercooled water falling aloft is insufficient to activate the ice pit, which is in fact a common mechanism for freezing rain in the southeast of china.

In the weather forecast business, the forecast of the freezing rain mainly depends on the diagnosis and forecast result of a numerical mode, and the subjective experience judgment of a forecaster is assisted. The weather forecast numerical mode is mainly realized by two methods, namely a display method and an implicit method, for the diagnosis and forecast of precipitation phase states including freezing rain. The explicit method adopts numerical simulation of the micro-physical process of the hydrogel, and predicts various near-surface rainfall by using the mixing ratio and the temperature of the hydrogel (Theriault et al 2010; Ikeda et al 2013); the implicit method mainly uses environmental characteristics of temperature and/or humidity to infer precipitation type (e.g., Baldwin et al 1994; Bourgouin 2000; Schuur et al 2012; Elmore and grades 2015; Chenard et al 2015), and because of the small spatial dimensions of the freezing rain process, it is difficult to show key processes that determine freezing rain formation, such as simulated hydrogel downstream transport and hydrogel interaction, even in convection-resolved high-resolution numerical weather forecast models. Therefore, implicit methods play a dominant role in the diagnostic prognostics of numerical patterns. Common implicit methods include Ramer algorithm, Baldwin algorithm, Bourgouin algorithm, NSSL algorithm, etc. (Ramer 1993, Baldwin1994, Bourgouin2000, Schuur 2012). Numerous studies have shown that rainfall phase prediction based on implicit methods has high prediction skills for snow and rain, but performs poorly in predicting freezing rain (Reeves2016, Elmore 2015).

In the current weather forecast business, the forecast of the precipitation phase in winter is based on a numerical mode, and manual correction is mainly performed by a forecaster, so that the requirements on accuracy and timeliness cannot be met. In recent years, a data-driven method based on machine learning can accurately predict with less dependence on physical details, and provides a new perspective and method for more efficiently and accurately predicting freezing rain. In the forecast work of rainfall phase introduced by machine learning method suitable for solving high dimensional non-linear problem in recent years, artificial intelligence methods such as decision tree, support vector machine and deep neural network are used by meteorological departments and scholars at home and abroad for exploring the forecast of the rainfall phase in winter, Reeves (2016) makes a decision tree to identify the mixed rainfall phase such as rain, snow, rain and snow, freezing rain and ice particles; on the basis of numerical prediction products, Dongquan et al (2013) develop an objective prediction model and products of the rain and snow phase state in the Chinese area by using an artificial neural network method, and can accurately predict the boundary of rain and snow in the north; the chrysol and the like (2021) establish a high-resolution objective classification model of the precipitation phase based on three machine learning methods XGboost, SVM and DNN respectively, and carry out comparison and inspection on the forecast effects of the three machine learning methods on the main precipitation phases of rain, sleet and snow 3 Jingjin wings under the same condition, thereby further improving the objective classification forecast skill of the complicated precipitation phase of sleet and snow. However, most methods mainly use near-earth meteorological elements or comprehensive index type elements in feature selection, and cannot well represent the vertical structures of atmospheric heat, water vapor and other elements which play a key role in the formation of precipitation phase; moreover, most methods are less (or neglected) for freezing rain forecasting and the forecasting effect is poor.

The characteristics input by the current precipitation phase artificial intelligence forecasting method based on machine learning are mostly limited to common variables such as temperature, humidity and wet bulb temperature, the information expression of high-altitude thermal power and water vapor structure is not complete (cellosolve, 2021; Reeves, 2016), the precipitation phase forecasting accuracy cannot be effectively improved, and particularly the forecasting capability of extreme precipitation events such as freezing rain is very low. In order to solve the problems that the physical cause and the dynamic mechanism of the freezing rain are complex and the physical model, the dynamic model and the linear model cannot be used for effective forecasting, the freezing rain needs to be forecasted by using an artificial intelligence technology.

However, the invention is an objective forecasting method for winter freezing rain forecast based on artificial intelligence technology and big data application. Compared with the existing method, the forecasting method has the advantages that the vertical section time sequence of variables such as atmospheric temperature, humidity, wind speed and vertical motion is adopted as the forecasting characteristic, the implicit forecasting model is constructed to carry out high-dimensional approximation on a physical mechanism formed by precipitation phase states including freezing rain, and the forecasting error is reduced by using basic meteorological variables with high observation and simulation precision instead of diagnostic indexes with complex error structures. Aiming at the problem of sample unbalance caused by small proportion of freezing rain in precipitation phase distribution, a new loss function is provided in the model training process, and the forecasting capacity of the algorithm on freezing rain is improved.

Disclosure of Invention

In view of the above technical problems in the related art, the present invention provides a method for forecasting freezing rain using artificial intelligence, which can overcome the above disadvantages of the prior art.

In order to achieve the technical purpose, the technical scheme of the invention is realized as follows:

a method for forecasting freezing rain by using artificial intelligence comprises the following steps:

s1, preprocessing the precipitation type observation data, and performing precipitation phase change according to the coding rule of the precipitation type record;

s2, randomly dividing the pretreated precipitation type observation data into a training data set and a verification data set according to the proportion of 90% to 10%;

s3, extracting features, selecting the temperature, the relative humidity, the wind speed, the vertical speed and the gravity potential of each atmospheric pressure layer in the high-altitude vertical dimension of the 500hPa-1000hPa atmosphere and the longitude and latitude and the height of each forecast station, performing geographic position matching on the reanalysis data 24 hours before the observation time according to the observation data time, and calculating according to the longitude and latitude information of the stations to obtain grid data of the reanalysis data;

s4, acquiring the temperature, the relative humidity, the wind speed in the north-south horizontal direction, the wind speed in the east-west vertical direction, the vertical speed and the gravity potential of all atmospheric pressure layers of 500hPa-1000hPa in the matched data, and analyzing the longitude, latitude and height data of each forecast station to construct a characteristic data set;

s5, training a forecasting model, namely training a rainfall phase forecasting model by using the observation data of the training data set as labels through a LightGBM algorithm to finally obtain and store an available freezing rain forecasting model;

and S6, directly forecasting to obtain a freezing rain precipitation phase state forecasting result by using the temperature, the relative humidity, the wind speed in the north-south horizontal direction, the wind speed in the east-west vertical direction, the vertical speed, the gravity potential and the longitude and the latitude and the height of each forecasting station as the input of the forecasting model in the last 24 hours at the current time.

Further, the precipitation phase-state conversion is specifically to be converted into liquid precipitation, solid precipitation, mixed phase precipitation and freezing rain phase precipitation.

Further, the grid data are calculated according to the longitude and latitude data information of the current station to obtain the nearest grid points in the reanalysis data, and then grid point data in a certain range around the grid points are selected.

Further, the forecasting result is a classification result of the four classification models, and the categories respectively correspond to rain, snow and sleet.

Further, the wind speed of the extraction barosphere includes the wind speed in the north-south horizontal direction and the wind speed in the east-west vertical direction, namely, the downwind and the transcendental wind.

The invention has the beneficial effects that: by adopting a complete characteristic engineering method and an efficient artificial intelligence algorithm framework, performing characteristic engineering processing on the temperature, the relative humidity, the wind speed in the north-south horizontal direction, the wind speed in the east-west vertical direction, the vertical speed and the gravity potential of all atmospheric pressure layers of 500hPa-1000hPa and the longitude, the latitude and the height of each forecasting station, and using the preprocessed observation data as tags, the rapid and accurate forecasting of the freezing rain in winter is realized.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without creative efforts.

Fig. 1 is a schematic diagram of precipitation phase information representation and data collection by using a method for forecasting freezing rain by artificial intelligence according to an embodiment of the invention.

Fig. 2 is a freezing rain prediction error test graph of a method for predicting freezing rain using artificial intelligence according to an embodiment of the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments of the present invention belong to the protection scope of the present invention, and for the convenience of understanding the above technical solutions of the present invention, the above technical solutions of the present invention are described in detail below by specific use modes.

The method for forecasting the freezing rain by utilizing the artificial intelligence comprises the following steps: first, for example: by using the freezing rain forecasting method, the winter precipitation type observation data of 2515 meteorological observation stations 2016 and 2018 in China for three winter and half years (10 months-3 months next year) are preprocessed. And the precipitation phase state is changed into rain, rain and snow, snow and freezing rain according to the coding rule of the precipitation type record (the precipitation types are various, and the application focuses on forecasting the freezing rain, so that the freezing rain precipitation types are distinguished). And randomly dividing the pretreated precipitation type observation data into a training data set and a verification data set according to the proportion of 90% to 10%.

The existing research shows that the change of the atmospheric vertical temperature profile is the main determining factor of the precipitation phase state, but is also influenced by other factors such as the precipitation particle phase state, the particle falling speed, the vertical movement, the terrain and the like, and because the temperature, humidity and wind speed data of a station and the periphery are difficult to obtain under the common condition, the invention adopts European center reanalysis ERA5 data which assimilates the observed data as a false true value to carry out characteristic engineering construction. And performing geographical position matching on the reanalysis data at the observation time every three hours according to the observation data time, and acquiring the temperature, the relative humidity, the south horizontal direction wind speed (latitudinal wind), the east-west vertical direction wind speed (latitudinal wind), the vertical speed, the gravity potential and the longitude, the latitude and the height of each forecast site to represent topographic information to construct a characteristic data set, wherein the temperature, the relative humidity and the north horizontal direction wind speed of 16 barolayers of 500hPa-1000hPa (comprising 500hPa, 550hPa, 600hPa, 650hPa, 700hPa, 750hPa, 775hPa, 800hPa, 825hPa, 850hPa, 875hPa, 900hPa, 925hPa, 950hPa, 975hPa and 1000 hPa) in the reanalysis data after matching. As shown in figure 1, the rainfall phase forecast feature space constructed by the method is 2307 dimensions in total, and sufficient information representation and data collection for the rainfall phase formation state vehicle are achieved.

For forecasting model training, the LightGBM is adopted as a basic algorithm framework for model training and forecasting. The gradient Boosting Decision tree GBDT (gradient Boosting Decision Tree) is a machine learning model framework with better performance, the main idea is to use a weak classifier (Decision tree) to carry out iterative training to obtain an optimal model, and the model has the advantages of good training effect, difficulty in overfitting and the like. The light GBM (light Gradient Boosting machine) is an efficient implementation version of the GBDT algorithm, supports efficient parallel training, and has the advantages of higher training speed, lower memory consumption, higher accuracy, supporting distributed processing of mass data and the like.

And taking the observation data of the training data set as a label, combining the input of the characteristic data set parameters, and carrying out rainfall phase state forecasting model training by loading and using a LightGBM model algorithm to finally obtain and store an available forecasting model.

Wherein, the invention provides a forecasting algorithm for 0-12 hours relative to 0 time and every 3 hours, therefore, the invention comprises the following 5 forecasting algorithms. The specific parameter settings of each forecasting algorithm are as follows:

0 hour: {

"num_leaves": 347,

"max_depth": 20,

"learning_rate": 0.13757700241802978,

"n_estimators": 123,

"subsample_for_bin": 226,

"min_split_gain": 0.04088659782698967,

"min_child_weight": 0.2618087363348358,

"min_child_samples": 190,

"subsample": 0.7522282288261666,

"subsample_freq": 29,

"colsample_bytree": 0.3936729186301381,

"reg_alpha": 0.02692277056238257,

"reg_lambda": 0.7431751883441085}

3 hours: {

"num_leaves": 87,

"max_depth": 16,

"learning_rate": 0.28501857817008797,

"n_estimators": 141,

"subsample_for_bin": 104,

"min_split_gain": 0.044704861240112304,

"min_child_weight": 0.3831165307373335,

"min_child_samples": 859,

"subsample": 0.8753916228505754,

"subsample_freq": 9,

"colsample_bytree": 0.42996341376509895,

"reg_alpha": 0.08472324841351359,

"reg_lambda": 0.4651052748489807}

6 hours: {

"num_leaves": 267,

"max_depth": 11,

"learning_rate": 0.1377491413266653,

"n_estimators": 168,

"subsample_for_bin": 181,

"min_split_gain": 0.19546915963967137,

"min_child_weight": 0.23942160973063564,

"min_child_samples": 105,

"subsample": 0.8808733396703929,

"subsample_freq": 56,

"colsample_bytree": 0.5386212621060782,

"reg_alpha": 0.3654787861181418,

"reg_lambda": 0.34829328731249587}

9 hours: {

"num_leaves": 466,

"max_depth": 15,

"learning_rate": 0.2805206780795043,

"n_estimators": 105,

"subsample_for_bin": 61,

"min_split_gain": 0.10013239820687801,

"min_child_weight": 0.21218388097084276,

"min_child_samples": 434,

"subsample": 0.06842995367405191,

"subsample_freq": 0,

"colsample_bytree": 0.9171552920925544,

"reg_alpha": 0.8974235565037368,

"reg_lambda": 0.35886080571998036}

12 hours: {

"num_leaves": 499,

"max_depth": 13,

"learning_rate": 0.23713595138960672,

"n_estimators": 140,

"subsample_for_bin": 104,

"min_split_gain": 0.007556084240058894,

"min_child_weight": 0.5893312731171183,

"min_child_samples": 138,

"subsample": 0.9399360127974601,

"subsample_freq": 36,

"colsample_bytree": 0.2855015410324987,

"reg_alpha": 0.10272211842823623,

"reg_lambda": 0.9813206832344459}

And finally, directly outputting and forecasting results of rainfall phase states after 0 hour, 3 hours, 6 hours, 9 hours and 12 hours by utilizing the characteristic data set of the forecasting model 24 hours before the current time. The forecasting result is a classification result of the four-classification model, the classification corresponds to rain, snow and sleet respectively, the output four classified input samples need to be concerned, the forecasting result is sleet, and the other rainfall phase state forecasting results are not sleet.

In addition, in order to improve the freezing rain forecast performance, the invention adopts an equalization mode for different types of weights in terms of the structure of the loss function, specifically, the invention automatically adjusts the weights by using the class frequency in the input data, so that the weights are in inverse proportion to the class frequency in the input data.

Numerical experiments of the ERA5 reanalysis data and the numerical model EC forecast data based on the European middle-term weather forecast center show that the accuracy of the forecast algorithm is higher, the forecast algorithm is embodied on a test set, and the TS scores of the method for forecasting the freezing rain at different timeliness are respectively as follows: 0.443 (0 hour), 0.473 (3 hours), 0.4497 (6 hours), 0.4461 (9 hours), 0.4255 (12 hours). Compared with a rainfall phase forecast product PTYPE of the EC, the TS score of the rainfall phase forecast product PTYPE of the EC for forecasting freezing rain at different aging times of 0 hour, 3 hours, 6 hours, 9 hours and 12 hours is between 0.12 and 0.2. The data was reanalyzed against ERA5 and the freezing rain TS score was 0.1421. For extreme weather conditions such as freezing rain to be forecasted, the TS score exceeds the result of ERA5 and EC by more than one time; the error sequence of the forecasting algorithm is stable, and the error increasing trend does not exist, as shown in figure 2; the forecasting algorithm has very high computational efficiency, and the average time for the model to complete inference on 507917 pieces of data is 30 seconds on a server with 2 Intel Xeon Gold 6240 CPUs (18 core 36 threads).

In summary, by means of the technical scheme of the invention, a complete characteristic engineering method and an efficient artificial intelligence algorithm framework are adopted, characteristic engineering processing is carried out on the temperature, the relative humidity, the wind speed in the north-south horizontal direction (latitudinal wind), the wind speed in the east-west vertical direction (latitudinal wind), the vertical speed and the gravity potential of 16 total atmospheric pressure layers of 500hPa-1000hPa, and the longitude, the latitude and the height of each forecasting station, and the preprocessed observation data are used as tags, so that the rapid and accurate forecasting of the freezing rain in winter is realized.

The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.

Claims

1. A method for forecasting freezing rain by using artificial intelligence is characterized by comprising the following steps:

2. The method for forecasting the freezing rain using artificial intelligence as claimed in claim 1, wherein the precipitation phase transition is specifically to be transformed into four categories of liquid precipitation, solid precipitation, mixed phase precipitation and freezing rain phase precipitation.

3. The method according to claim 1, wherein the grid data is the most adjacent grid points in the reanalysis data calculated according to the longitude and latitude data information of the current station, and a certain range of grid point data around the grid points is selected.

4. The method for forecasting freezing rain using artificial intelligence according to claim 1, wherein the forecasting result is a classification result of four classification models, and the classification corresponds to rain, sleet, snow and freezing rain.

5. The method for forecasting freezing rain using artificial intelligence according to claim 1, wherein the wind speeds of the extracted barosphere include north and south horizontal wind speeds and east and west vertical wind speeds, i.e., both the mainwind and the transcendental wind.