| Tropical cyclones are one of the major catastrophic weather systems affecting China.Improving the level of research and forecast of tropical cyclones is of great significance to economic development and national defense construction.In recent years,the objective prediction level of tropical cyclone paths has been significantly improved,but its intensity prediction has improved to a lesser extent.The uncertainty of long-term change prediction such as generation frequency is still very large,and it is still the focus of scholars.The physical mechanism that affects the frequency of tropical cyclones is complex.In August 2018,a total of 9 tropical cyclones were generated,which is significantly larger than the number of climatic states.This article uses statistical analysis methods to analyze the physical factors that have a significant impact.The results show that the relative positive vorticity of the 850 h Pa relative vorticity and the multi-year average in August 2018 reached 0.25×10-4/s,and the relative humidity at 600 h Pa was 8%-10%higher than the multi-year average,which became the main reason for the abnormality.This is related to the location characteristics of the Northwest Pacific subtropical high,the anomalous westerly winds from the North Indian Ocean to the South China Sea,the meridional wind anomalies in the South China Sea,the anomalous water vapor flux,and the location of the monsoon trough.In this paper,through statistical analysis of tropical cyclones with intensity mutations from 1979 to 2018,it is found that the areas where tropical cyclones are most likely to suddenly increase are 15°-20°N and125°-135°E in the east of the Philippines.The tropical cyclones that suddenly increased in this area were selected.For the first time,the hourly resolution sea temperature data was used,and the method of sea temperature area accumulation was used.It was found that the sea temperature data had more advantages in the study of the effect of intensity,and the intensity correlation coefficient over 0.94.Regional accumulation is a more reasonable method in the study of the relationship between sea temperature and intensity.Compared with the central instantaneous sea surface temperature value and the central wind speed,the regional cumulative sea surface temperature value has a better correlation,and the correlation coefficient reaches above 0.8,which can be applied to the analysis of other physical factors.Based on the NCEP/NR1 reanalysis data from the United States,the atmospheric factors,ocean factors,underlying surface factors,and the cyclone's own position,speed and other physical factors are extracted from the tropical cyclone passing area,and the changes of the above factors in the past 12 hours are taken as variables,multiple linear regression method for establishing intensity of tropical cyclones statistical forecasting model,select United States NECP/GEFS forecast data,maximum wind speed variation of the intensity of tropical cyclones in 2016 all the intensity and stronger cyclones and predict the next 12hours center,and with GEFS forecast results compared.The statistical model's prediction effect on the maximum wind speed of the tropical cyclone center is better than the GEFS result,especially the prediction of extreme values is more accurate.When the correlation coefficient exceeds 0.5,the root-mean-square error can be reduced by a maximum of7-8m/s.Using the random forest algorithm,the same data are used to learn the tropical cyclone intensity forecast,to examine the applicability of the machine learning method,and try to improve the numerical forecast results.It is found that the machine learning method of random forest has a better ability to predict the intensity of tropical cyclones,which is improved from the GEFS results.The random forest machine learning method mostly keeps the root-mean-square error of tropical cyclone forecast in the next 12 hours at 6-7m/s,and the correlation coefficient reaches 0.6.The correlation coefficient of the prediction results of wind speed is 0.5-0.6,and the root mean square error can be reduced by 6-7m/s at most.In order to compare the forecasting effects of machine learning methods and statistical methods,tropical cyclone intensity from 2015 to2016 was selected as sample data to forecast the 24-hour,48-hour and72-hour intensity of tropical cyclone in 2018.The results show that both the statistical model and machine learning have a better effect on the prediction results of the wind speed variables of 24,48 and 72 hours in2018 than the GEFS prediction results,and the machine learning effect is better than the regression fitting effect of the statistical model,but the final prediction error is slightly smaller than that of the statistical model.After integrating machine learning methods and statistical models,the results of 24-hour and 72-hour forecasts are improved more obviously.In the 24-hour forecast of wind speed variable,the optimal forecast result is 44.6%higher than the GEFS forecast result.In the 48-hour wind speed forecast,the optimal forecast result is 31.9%higher than the GEFS forecast result.In the 72-hour wind speed forecast,the optimal forecast result is 26.9%higher than the GEFS forecast result.Analysis found that the radius,the initial wind speed,the vorticity of 500 h Pa and the divergence of 200 h Pa,relative humidity,and SST are highest factor to affect the intensity of tropical cyclone,these factors affect the mass transport within and outside the cyclone by affecting the maintenance of cumulonimbus clouds and the release of heat,ultimately affect the intensity of tropical cyclones.The forecast error is mainly caused by the position deviation of tropical cyclone forecast,the value of influence factor and the spatial distribution of forecast error. |
PDF Full Text Request